Mimecast Warns Five in Six Enterprise AI Agents Lack Basic Safeguards

From Shadow to Spotlight: Building an AI Governance Strategy With Mimecast North America — Ai4

Ai4

Banning generative AI backfires by accelerating shadow tool adoption, requiring firms to deploy adaptive source-and-destination guardrails as autonomous agents multiply insider risk around the clock.

AI-assisted summary of From Shadow to Spotlight: Building an AI Governance Strategy With Mimecast North America — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 17, 2026, this presentation features Giulian Garruba, Senior Director of Product Marketing at Mimecast North America. The session examines how unsanctioned AI tools and autonomous agents create unmonitored pathways for enterprise data exposure, offering strategies for security leaders to achieve visibility, build non-restrictive governance frameworks, and prevent intellectual property leaks.

## TL;DR Giulian Garruba outlines the compounding corporate risks driven by Shadow AI and the rapid emergence of autonomous agentic systems operating at machine speed. Highlighting telemetry that reveals massive hourly transfers of proprietary source code and unsanctioned tool adoption even among security personnel, he demonstrates why blanket tool bans inevitably fail. To regain control, Garruba urges CISOs to answer four foundational visibility questions regarding tools, agents, connectors, and user intent. He proposes replacing blunt restrictions with an adaptive, 'well-lit path' that leverages dual-vector filtering, behavioral nudges, and harmonized compliance across the EU AI Act, NIST AI RMF, and ISO/IEC 42001.

## Atmosphere & style The presentation carries an urgent, pragmatic, and analytical tone, unfolding like an executive briefing designed to shake security leaders out of complacency. Garruba balances sobering telemetry and cautionary real-world incidents with accessible analogies like automotive seatbelts and engineered dams. The atmosphere feels authoritative and measured, shifting the narrative arc from mounting agentic vulnerability toward actionable, auditable governance.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=4s) Enterprise insider risk has evolved from traditional Shadow IT to Shadow AI, and is now leaping into autonomous agentic systems operating as independent decision-makers. - [01:50](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=110s) The historical technology governance cycle—moving from exploration to risk discovery to governance enforcement—has dramatically compressed for AI compared to previous technological revolutions. - [03:26](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=206s) Shadow AI is widespread: 80% of general enterprise users and 90% of InfoSec personnel admit to using unsanctioned AI tools, with 57% inputting sensitive corporate data into public models. - [03:26](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=206s) Any corporate insider risk, data protection, or AI policy not reviewed within the past 90 days should be considered outdated. - [05:03](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=303s) While over one billion AI agents are projected to be deployed by 2029, only 1 in 6 currently operating in Fortune 500 companies is properly secured. - [05:03](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=303s) Enterprise insider risk is non-uniform, with approximately 8% of users generating 80% of all security incidents—a dynamic intensified by expanding AI access. - [07:14](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=434s) Proprietary source code represents 43% of all content submitted to external AI platforms, exposing core intellectual property at high velocity. - [08:05](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=485s) Agentic failures frequently occur within legitimate technical permissions rather than through perimeter breaches, executing unintended decisions at machine speed. - [09:12](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=552s) Autonomous agents act as the 'ultimate insiders' because they hold legitimate credentials, understand corporate networks, and operate continuously 24/7 without biological limitations. - [11:08](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=668s) To close the visibility gap, CISOs must baseline four areas: active tool inventory, agent identity tracking, connector/MCP server permissions, and the underlying user intent driving workarounds. - [12:27](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=747s) Outright bans on generative AI are ineffective and accelerate shadow usage; security teams should instead create a 'well-lit path' using risk-calibrated, adaptive source-and-destination controls. - [15:10](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=910s) A complete AI compliance posture harmonizes the EU AI Act for legal mandates, the NIST AI RMF for operational methodology, and ISO/IEC 42001 for certifiable management systems. - [17:02](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=1022s) Generative AI democratizes machine-speed execution across the broader 92% of the workforce, significantly increasing the potential blast radius of inadvertent data exposure.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=4s) From Shadow IT to Shadow AI and Agentic Risk

Giulian Garruba, Senior Director of Product Marketing at Mimecast, introduces the session by highlighting how artificial intelligence has rapidly evolved from an office curiosity into an urgent, board-level conversation. The title of the session, *From Shadow to Spotlight*, reflects the reality of unsanctioned tools operating quietly across corporate networks. Garruba explains that the classic challenge of insider risk has evolved from traditional shadow IT into "Shadow AI," and is now taking a profound leap forward with autonomous agentic systems.

To tackle this mounting risk, organizations must transition from reactive defense to proactive visibility. Garruba outlines a three-part agenda for the presentation: examining the fundamental transformation driven by shadow AI and autonomous agents, establishing an actionable AI governance framework, and deploying proactive, adaptive security to regain visibility and control over enterprise data.

```mermaid flowchart LR accTitle: Evolution of Enterprise Insider Threat subgraph Past["Legacy IT"] A["Shadow IT<br/>(Unsanctioned Apps)"] end subgraph Present["Generative AI"] B["Shadow AI<br/>(Prompts & Data Sharing)"] end subgraph Future["Autonomous Scale"] C["Agentic AI<br/>(Autonomous Decision-Makers)"] end A --> B --> C ```

---

[01:50](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=110s) The Shrinking Evolution of Tech Governance and the Seatbelt Analogy

Every major technological transformation radically alters how society functions and how organizations conduct business. Garruba explains that historically, technology revolutions pass through three distinct phases: broad exploration, risk discovery, and governance enforcement.

To illustrate this dynamic, he draws an analogy to the automotive industry. In the early exploration phase, society transitioned from horse-drawn carriages to mainstream motor vehicles. The subsequent "risk discovery" phase revealed a harsh reality: while buggy crashes resulted in relatively low damage, automobile collisions caused severe injuries as passengers were violently thrown from vehicles. The eventual governance response was the invention and mandatory enforcement of seatbelts. However, Garruba stresses that the timeline across these three phases has dramatically collapsed for AI; modern security and compliance teams no longer have decades to iterate, as shifts in the AI threat landscape occur faster than ever before.

---

[03:26](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=206s) The Reality of Shadow AI in Enterprise Environments

Generative AI is no longer a niche technology; it has become universally mainstream, experiencing a 9x explosion in enterprise usage between 2024 and 2025. Garruba notes that while roughly 80% of organizations openly acknowledge using GenAI, the remaining 20% that claim not to use it simply suffer from a lack of internal visibility.

Despite the proliferation of AI, safe enterprise usage is rarely governed by well-defined guardrails. Garruba shares alarming industry data showing that an overwhelming percentage of employees actively bypass organizational policies. Crucially, any insider risk, data protection, or AI policy that has not been re-evaluated within the past 90 days should already be considered outdated.

- **80% of general enterprise users** admit to using unapproved "shadow AI" tools without official organizational permission. - **90% of InfoSec personnel** admit to using unsanctioned AI tools in their day-to-day work. - **57% of users** admit to entering sensitive corporate data—such as passwords, internal source code, and financial information—directly into public AI models.

---

[05:03](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=303s) The Agentic Multiplier and Non-Uniform Human Risk

While shadow AI represents the human dimension of risk, autonomous agentic AI significantly multiplies that exposure. Citing projections from IDC, Garruba notes that organizations will deploy over one billion AI agents by 2029. Unlike standard generative AI chatbots that rely on simple call-and-response interactions, AI agents act as independent decision-makers. In practice, organizations assign them complex mandates—effectively saying, "Here's a problem. Solve it by Tuesday. Thank you"—without conducting the background checks, vetting, or verification traditionally applied to human hires.

At the same time, human organizational risk is unevenly distributed. Mimecast’s research reveals that approximately 8% of enterprise users generate 80% of all security incidents. In an era marked by corporate consolidation, workforce reductions, and mandates to "do more with AI," individual human users wield greater access and responsibility than ever before. This 8% typically comprises senior leaders with broad financial or administrative privileges, careless staff who do not grasp the implications of their actions, or departing employees seeking to exfiltrate proprietary data. To understand AI risk, Garruba argues that security teams must fundamentally start by analyzing human behavior.

```mermaid flowchart TD accTitle: Disconnect in Workforce Governance subgraph HumanHiring["Human Employee Vetting"] H1["Resume Review"] --> H2["Background Checks"] H2 --> H3["Access Governance"] end subgraph AgenticDeployment["AI Agent Deployment"] A1["Task Assignment:<br/>'Solve by Tuesday'"] --> A2["Autonomous Execution"] A2 --> A3["Unsecured Access<br/>(5 in 6 Fortune 500 Unsecured)"] end ```

- **Only 1 in 6 AI agents** currently deployed across Fortune 500 companies is properly secured. - Individual users currently orchestrating two or three agents are projected to manage thousands of autonomous agents in the near future. - **8% of users** generate **80% of organizational security incidents**, proving that insider risk is non-uniform.

---

[07:14](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=434s) Telemetry from the Front Lines: High-Volume Code and File Leaks

Telemetry gathered from Mimecast’s AI monitoring and insider risk platforms illustrates the sheer velocity of data being pushed into external AI applications. Every single hour, enterprise environments log massive volumes of data transfer actions, amounting to roughly 2.9 million total interactions per month.

Garruba identifies the heavy flow of proprietary code as the most alarming trend. Because software developers and engineers are under relentless pressure to ship features rapidly, they actively feed core intellectual property into public AI models. He emphasizes to the audience that these data transfers are unquestionably happening within their companies; the only real question is whether the security team has the telemetry required to see it.

- **5,500+ copy-and-paste events** into external AI tools recorded every hour. - **2,700+ direct file uploads** to AI platforms executed every hour. - **43% of all content** submitted to external AI tools consists of proprietary source code.

---

[08:05](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=485s) Case Study: Autonomous Agent Failure at Meta

To ground agentic risk in a concrete scenario, Garruba recounts an incident that occurred early in 2026 at Meta. An engineer submitted a sensitive question to an internal technical forum intended strictly for a limited group of engineers. An internal AI agent assigned to answer the question fundamentally misinterpreted the request's context and permissions, reposting the sensitive technical inquiry to a widely accessible internal forum seen by thousands of unauthorized employees.

Meta classified the incident at its highest severity level for critical internal exposure, even though the data remained exposed for only two hours. Garruba stresses the core lesson of this incident: the AI agent did not breach a perimeter, bypass a firewall, or violate technical rule sets. Instead, it operated strictly within its granted technical access, but executed an autonomous decision at machine speed outside its intended operational boundaries.

---

[09:12](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=552s) "The Call Is Coming from Inside the House": Agents as Ultimate Insiders

Garruba warns that AI agents will rapidly become the "ultimate insiders" within modern organizations. Because agents possess legitimate access credentials, understand internal systems, and map out the corporate directory, the classic threat metaphor applies: "The call is coming from inside the house." Unlike human workers, autonomous agents do not get tired, do not second-guess their choices, operate at exponential scale, and execute decisions at machine speed.

```mermaid flowchart TD accTitle: Comparing Human vs Agentic Insider Risk subgraph HumanRisk["Human Insider Risk"] H_Limit["Active between keyboard & fingers"] H_Sleep["Dormant overnight/weekends"] H_Speed["Human cognitive speed"] end subgraph AgenticRisk["Agentic Insider Risk"] A_Limit["Unbounded autonomous permissions"] A_Sleep["Continuous 24/7/365 execution"] A_Speed["Exponential machine speed"] end ```

In traditional security environments, human risk is concentrated "between the fingers and the keyboard"—meaning that when employees log off to go to sleep, incident creation drops off until the next morning's emails, messages, and downloads resume. AI agents, by contrast, run continuously 24/7. Rather than managing fewer insider variables, enterprise security teams now face thousands of new, autonomous insiders that possess expanded permissions but zero biological limitations.

Transitioning into the solution space, Garruba asserts that managing this new surface requires a three-pillar governance framework: **Visibility, Control, and Governance**. He turns to the audience of security and compliance leaders to poll how many organizations currently maintain and enforce an active AI Acceptable Use Policy.

[11:08](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=668s) The Visibility Gap: Four Critical Questions for CISOs

Polling the room on organizational readiness reveals a stark reality: while a handful of hands stay up when asked who has controls in place to track sanctioned versus unsanctioned AI tools and their proper usage, almost every hand drops when asked who could produce detailed audit logs of daily employee AI chats and data flows during an audit. With only a single hand remaining in the audience, the gap between current security postures and actual visibility is massive. To bridge this divide, security leaders must begin not with rigid blockades, but with baseline visibility structured around four core questions that every CISO must be able to answer.

```mermaid flowchart TD accTitle: Four Core Visibility Questions for AI Governance subgraph VisibilityQuestions["Core Visibility Questions for CISOs"] Q1["1. Tool Inventory<br/>Which GenAI tools are in active use?"] Q2["2. Agent Tracking<br/>Which agents are running & who runs them?"] Q3["3. Connector Access<br/>What access do MCP servers & 3rd-party agents have?"] Q4["4. User Intent<br/>Why are employees circumventing policy?"] end

Q1 --> Flow["Establish Baseline Visibility"] Q2 --> Flow Q3 --> Flow Q4 --> Flow ```

The first question is identifying which generative AI tools are actively in use across the enterprise. The second is discovering which AI agents are executing and who deployed them. The third involves mapping the access granted to AI connectors, such as Model Context Protocol (MCP) servers and third-party commercial agents embedded in enterprise software. Finally, CISOs must understand user intent—the underlying "why" behind the adoption of shadow tools. Circumvention almost always occurs because legitimate business needs are unmet; when an enterprise offers only a single approved model and an employee cannot generate the specific graphic or answer they need, they immediately pivot to an unvetted alternative, instantly creating insider risk.

- **Tool Inventory**: Identifying every generative AI application actively touched by employees across departments. - **Agent Tracking**: Pinpointing autonomous and semi-autonomous agents operating within the network and the specific identities operating them. - **Connector Auditing**: Evaluating permissions, data pathways, and integration hooks for MCP servers and vendor-provided AI connectors. - **Intent Analysis**: Diagnosing functional gaps in approved toolsets that drive employees toward unsanctioned workarounds.

---

[12:27](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=747s) Adaptive Controls, Agent Guidance, and the Well-Lit Path

Treating enterprise AI security as an on-off switch is a critical operational mistake. Banning tools outright does not solve the problem; instead, security teams must deploy adaptive controls that reflect the continuous spectrum of risk from low to critical. A marketing team drafting a public-facing blog post citing publicly available data carries minimal exposure, whereas an autonomous agent accessing confidential, unreleased financial metrics for an upcoming earnings report represents severe risk. Effective protection requires combining "block by source" (preventing sensitive databases, proprietary codebases, or financial records from ever flowing to AI systems) with "block by destination" (restricting traffic strictly to authorized AI platforms), while reserving aggressive measures like endpoint quarantines and credential revocation for true emergencies.

```mermaid flowchart LR accTitle: Adaptive AI Control and Governance Flow subgraph Ingestion["Source Data"] S1["Public Data"] S2["Sensitive Financials / Code"] end

subgraph PolicyEngine["Adaptive Security Controls"] BlockSrc["Block by Source"] BlockDst["Block by Destination"] Nudge["Slack / In-App Nudges"] end

subgraph Destinations["Execution Layer"] ApprovedAI["Sanctioned AI / Agents"] BlockedAI["Unsanctioned AI / Leakage"] end

S1 --> ApprovedAI S2 --> BlockSrc BlockSrc -.->|Prevent Exfiltration| BlockedAI ApprovedAI --> Nudge ```

Governance must also weave education directly into workflows through real-time interventions, such as behavioral nudges over Slack alerting employees when they interact with unapproved tools. Crucially, education and guardrail enforcement must soon extend beyond human workers to the AI agents themselves, guiding and restricting autonomous actors that frequently wander outside prompt boundaries. Drawing on an insight from the Google Cloud Office of the CISO, "blocking generative AI rather than creating a well-lit path for organizations to use it and securely is ineffective. It often precipitates the proliferation of shadow and data leakage—the very risk enterprises were trying to avoid." Much like building a dam with engineered relief valves rather than trying to hold back a rushing creek with bare hands, organizations must channel employee momentum so that "the safe path is the easy path."

- **Risk-Calibrated Controls**: Differentiating between low-risk tasks (summarizing public documents) and high-stakes operations (analyzing proprietary financials). - **Dual-Vector Filtering**: Enforcing boundaries simultaneously at the data source level and at the external AI destination level. - **Agent and User Nudging**: Utilizing automated chat-based alerts to redirect human behavior while establishing training guardrails for autonomous agents. - **Relief-Valve Strategy**: Providing sanctioned, low-friction AI alternatives rather than strict bans that inevitably spawn underground Shadow AI workarounds.

---

[15:10](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=910s) AI Governance Frameworks and Regulatory Compliance

Comprehensive governance requires demonstrating three core capabilities to boards, external auditors, and regulatory bodies: drastically reducing the attack surface of unsanctioned AI, quantifying and managing the combined risk of human operators and autonomous AI agents, and maintaining full data provenance across every communication channel. If an AI agent breaks past boundaries and exfiltrates proprietary records, the organization must have the granular telemetry required to pinpoint the exact prompt or chat submitted to Microsoft Copilot or Anthropic Claude, identify who initiated the session, and reconstruct the exact sequence of events.

```mermaid flowchart TD accTitle: Three-Tier AI Compliance and Governance Model subgraph Standards["Comprehensive AI Compliance Model"] EUA["EU AI Act<br/>(Mandatory Legal Requirements)"] NIST["NIST AI RMF<br/>(Methodology & Best Practices)"] ISO["ISO/IEC 42001<br/>(Certifiable Management Standard)"] end

subgraph Execution["Enterprise AI Governance"] Gov["Audit Readiness, Risk Mitigation & Telemetry"] end

EUA --> Gov NIST --> Gov ISO --> Gov ```

This level of rigor is rapidly shifting from best practice to mandatory compliance under emerging global regulations. The enforcement of the EU AI Act applies not only to European enterprises but to any entity placing AI systems on the EU market, functioning with the broad extraterritorial reach of GDPR. Alongside regional laws, the NIST AI Risk Management Framework provides an actionable methodology rapidly becoming the de facto US standard, while ISO/IEC 42001 delivers the first auditable, certifiable global standard for AI management systems (with Mimecast serving as the first cybersecurity vendor to achieve certification). Organizations should harmonize these standards into a unified posture where the EU AI Act supplies legal compliance, NIST provides the operational methodology, and ISO/IEC 42001 establishes the auditable management system.

- **Auditable Prompts and Telemetry**: Capturing complete prompt records, agent configurations, and user attribution across enterprise interactions. - **EU AI Act Impact**: Navigating extraterritorial compliance obligations similar to GDPR for all organizations offering AI-driven services in European markets. - **NIST AI RMF**: Implementing government-backed voluntary risk management guidelines as an operational framework. - **ISO/IEC 42001 Standard**: Benchmarking internal AI security against the premier certifiable international standard for artificial intelligence governance.

---

[17:02](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=1022s) Managing the Speed of Insider Risk and Next Steps

The traditional insider threat dynamic—where 8% of users generate 80% of an organization's security risk—is undergoing a dangerous transformation. While that high-risk 8% is moving faster than ever, the remaining 92% of the workforce is now also accelerating exponentially, equipped with generative tools that allow them to generate code, draft communications, and execute tasks at machine speed. With AI democratizing technical execution, organizations must confront the reality of absorbing thousands of supercharged "new insiders" across their networks.

To help enterprises regain visibility into this rapid expansion, Mimecast offers a 30-day proof of value assessment. This evaluation maps out active users interacting with AI, tracks outbound data destinations, catalogues deployed autonomous agents, and breaks down the high-risk 8% user segment by individual name, department, and geographic location to establish immediate operational control over AI-driven insider risk before opening the floor to audience questions.

- **Machine-Speed Workforce**: Accelerated productivity across the broader 92% of employees amplifies the potential blast radius of inadvertent data exposure. - **Granular Risk Mapping**: Profiling high-risk AI usage patterns down to the specific user, team, and regional office. - **30-Day Assessment**: Rapidly baselining shadow AI applications, agent behaviors, and exfiltration paths across enterprise environments.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=4s) From Shadow to Spotlight — Giulian Garruba

> [05:03](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=303s) Here's a problem. Solve it by Tuesday. Thank you — Giulian Garruba

> [05:03](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=303s) do more with AI — Giulian Garruba

> [09:12](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=552s) The call is coming from inside the house. — Giulian Garruba

> [09:12](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=552s) between the fingers and the keyboard — Giulian Garruba

> [12:27](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=747s) block by source — Giulian Garruba

> [12:27](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=747s) block by destination — Giulian Garruba

> [12:27](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=747s) blocking generative AI rather than creating a well-lit path for organizations to use it and securely is ineffective. It often precipitates the proliferation of shadow and data leakage—the very risk enterprises were trying to avoid. — Google Cloud Office of the CISO

> [12:27](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=747s) the safe path is the easy path. — Giulian Garruba

## Mentioned Giulian Garruba, Mimecast, Ai4, IDC, Fortune 500, Meta, Model Context Protocol (MCP), Slack, Google Cloud Office of the CISO, Microsoft Copilot, Anthropic Claude, EU AI Act, GDPR, NIST AI Risk Management Framework, ISO/IEC 42001

## Video description Giulian Garruba - Senior Director, Product Marketing, Mimecast North America

AI is everywhere in the workplace, and so is Shadow AI. Unsanctioned tools and autonomous agents are quietly opening new paths for data exposure. This session covers moving from blind spots to visibility: understanding how employees interact with GenAI and AI agents, building governance frameworks that don't kill innovation, and preventing data leaks before they hit headlines. Get actionable guidance on how to accelerate productivity & innovation with AI without losing control of your data.

[Source](https://www.youtube.com/watch?v=BWtXPtvRlW8)

中文简报

Mimecast警告：六分之五的企业AI智能体缺乏基本防护 禁止生成式AI因加速影子工具的采用而适得其反，随着自主智能体全天候加剧内部风险，企业需要部署自适应的源与目的地防护机制。 Giulian Garruba阐述了由影子AI（Shadow AI）以及以机器速度运行的自主智能体系统快速涌现所带来的企业复合风险。他强调遥测数据显示每小时都有大量专有源代码被传输，且即便在安全人员中也存在使用未经批准工具的情况，并据此说明了为何一刀切式的工具禁令必然会失败。为了重获控制权，Garruba敦促各CISO回答关于工具、智能体、连接器以及用户意图的四个基础可见性问题。他提议用一条自适应的“明亮路径”取代生硬限制，利用双向过滤、行为助推，并结合EU AI Act、NIST AI RMF和ISO/IEC 42001实现协同合规。 企业内部风险已从传统的影子IT演变为影子AI，目前正迈向作为独立决策者运行的自主智能体系统。 (0:04) 与以往的技术革命相比，AI经历的“从探索到风险发现再到治理执行”的历史技术治理周期已被大幅压缩。 (1:50) 影子AI十分普遍：80%的企业普通用户和90%的InfoSec人员承认使用未经许可的AI工具，其中57%将企业敏感数据输入到了公共模型中。 (3:26) 任何在过去90天内未进行审查的企业内部风险、数据保护或AI策略，均应被视为已过时。 (3:26) 尽管预计到2029年将部署超过10亿个AI智能体，但目前在Fortune 500企业中运行的智能体中仅有六分之一得到了适当防护。 (5:03)

Reading summary

Giulian Garruba of Mimecast argues that enterprise security faces an unprecedented insider threat as unvetted autonomous AI agents and unsanctioned shadow AI operate at machine speed inside corporate networks. To manage this expanding attack surface, organizations must abandon ineffective blanket bans in favor of adaptive governance that couples granular telemetry with international compliance frameworks. The Escalation from Shadow AI to Autonomous Agents (0:04) Generative AI enterprise adoption experienced a ninefold surge between 2024 and 2025, dramatically outpacing internal security policy cycles. According to survey data presented by Garruba, 80% of general enterprise workers and 90% of information security personnel admit to using unsanctioned AI applications, with 57% of users confessing to feeding sensitive corporate data—such as passwords, financial records, and source code—directly into public AI models. Garruba emphasized that any corporate AI or insider risk policy not re-evaluated within the prior 90 days is already obsolete. This ungoverned environment is further complicated by the emergence of autonomous agentic systems. Citing IDC projections, Garruba noted that organizations are expected to deploy more than one billion AI agents by 2029. Unlike passive chatbots, these agents operate as independent decision-makers assigned broad mandates without the background checks, vetting, or identity governance standard for human employees. Currently, an estimated five out of six AI agents deployed across Fortune 500 enterprises remain unsecured. Autonomous Agents as Continuous Inside Threats (8:05) Because autonomous agents possess verified system credentials, navigate corporate directories, and hold authorized access to internal data, Garruba characterized them as the ultimate insider threats operating from within the organization. Unlike human employees whose operational risk is bounded by working hours, autonomous agents run continuously without fatigue and execute decisions at machine speed. To demonstrate this failure mode, Garruba cited an early 2026 incident at Meta where an internal AI agent was asked to answer a technical inquiry in a restricted developer forum. Operating entirely within its technical access rights, the agent misjudged the inquiry's context and reposted the sensitive technical question to a widely accessible forum viewed by thousands of unauthorized personnel. Garruba stressed that this high-severity breach occurred without perimeter violations or firewall exploits, proving that automated actors can cause severe exposure purely through unguided, authorized autonomy. Data Leakage Telemetry and Asymmetric Human Risk (7:14) Telemetry from Mimecast's monitoring platforms captures approximately 2.9 million monthly AI interactions in enterprise networks, recording over 5,500 copy-and-paste actions and 2,700 direct file uploads to external AI platforms every hour. Proprietary source code constitutes 43% of all content uploaded to these external models, driven largely by software engineers seeking to accelerate feature delivery. This continuous data transfer is shaped by an asymmetric risk profile in human behavior. Mimecast research indicates that roughly 8% of enterprise users—frequently privileged executives, careless employees, or departing staff—generate 80% of all organizational security incidents. However, Garruba warned that generative tools now equip the remaining 92% of workers to draft communications, write code, and move enterprise data at machine speed, significantly expanding the blast radius across the wider workforce. Adaptive Controls and Standardized Governance Frameworks (12:27) Addressing the visibility gap requires CISOs to answer four foundational questions: which AI tools are active, which agents are running, what access is granted to connectors like Model Context Protocol (MCP) servers, and why users bypass policy. Citing guidance from the Google Cloud Office of the CISO, Garruba warned that rigid bans inevitably provoke underground workarounds. Instead, security teams must design a 'well-lit path' using adaptive, risk-calibrated controls that combine source blocking for critical assets, destination filtering, and real-time behavioral nudges over communication channels like Slack. To satisfy boards and regulatory audits, organizations must demonstrate complete data provenance, prompt logging, and user attribution. Garruba recommended harmonizing three complementary structures into a single framework: the EU AI Act to address mandatory legal obligations across European markets, the NIST AI Risk Management Framework for operational methodology, and ISO/IEC 42001 to supply an auditable international management standard. Garruba concluded by presenting Mimecast's 30-day assessment to baseline shadow AI tools and profile high-risk user segments. However, an unresolved question remains whether enterprise security teams can realistically establish auditability and prompt-level provenance across thousands of rapidly evolving semi-autonomous connectors before agentic adoption outpaces centralized governance capabilities.

精炼阅读摘要

Mimecast产品营销高级总监Giulian Garruba指出，企业内部风险正从传统影子IT加速演变为影子AI与自主智能体风险；面对激增的数据外泄与合规压力，安全团队必须从被动封堵转向以可见性、自适应控制和审计合规为核心的治理框架。 影子AI蔓延与内部高危行为现状 (3:26) 生成式AI在企业中的应用呈现爆发式增长，Garruba引用的数据显示其企业使用量在2024至2025年间增长了9倍。虽然约80%的企业承认使用生成式AI，但另有80%的普通企业用户以及高达90%的信息安全人员承认在日常工作中使用未经批准的影子AI工具。更为严峻的是，57%的用户承认直接将内部源代码、密码和财务信息等敏感数据输入公共模型中，这使得任何超过90天未重新评估的内部风险或数据保护政策面临失效风险。 根据Mimecast的监控平台遥测，企业环境每月产生的外部AI交互约达2.9百万次，其中包括每小时超过5500次复制粘贴和2700多次直接文件上传。最关键的泄露隐患来自软件研发端：提交至外部AI工具的内容中有43%为专有源代码。与此同时，企业内部风险表现出高度不均衡性，研究表明约8%的企业用户产生了全公司80%的安全事件，而生成式工具正让这部分高特权、疏忽大意或即将离职的高危人员以更快速度造成数据暴露。 智能体风险乘数与终极内部人威胁 (5:03) Garruba指出，自主AI智能体的出现使安全暴露面呈现乘数级放大。IDC预测到2029年企业部署的AI智能体将超过10亿个。与仅依赖提示词问答的传统聊天机器人不同，智能体作为独立决策者被指派复杂任务，但企业往往未对其进行类似于人类员工的背景调查或资质审核。目前财富500强企业中部署的AI智能体仅有约六分之一具备充分的安全防护。 智能体正在演变成企业的“终极内部人”，它们拥有合法凭证，熟悉内部网络，且不受人类生理限制，能够以机器速度在全天候持续执行决策。Garruba列举了Meta在2026年初发生的事故：一名工程师在仅对特定团队开放的内部技术论坛提问，负责解答的内部AI智能体误判了权限与上下文，将高度敏感的技术问询直接公开发布至数千名未授权员工可见的通用论坛。该事件在两小时内被Meta定级为最高严重级别，印证了智能体无需越过防火墙或攻破边界，仅凭既有权限的自主决策即可触发重大安全暴露。 可见性缺口与自适应控制策略 (11:08) 现场调研显示，大部分企业管理者无法在审计中提供员工日常AI对话与数据流的详尽日志。为弥补可见性鸿沟，CISO必须明确回答四大核心问题：当前活跃使用的生成式AI工具有哪些、运行中的智能体及其操作者身份为何、模型上下文协议（MCP）服务器与第三方智能体拥有何种权限，以及员工绕过合规策略的真实意图。分析表明，员工规避规则多因官方工具无法满足工作需求，单纯的封禁往往会适得其反，加剧影子AI的地下蔓延。 Garruba主张采取“明亮路径”（well-lit path）和分级自适应控制，摒弃“全盘允许或全盘禁止”的粗暴模式。安全架构应实施“按数据源阻断”（禁止敏感数据库与专有代码流向AI）与“按目的地阻断”（仅放行受批准平台）的双向过滤。同时，企业应在工作流中引入轻量级引导，例如通过Slack发送实时提示以规范员工行为，并将安全护栏延伸至防止自主智能体越界操作，实现安全路径即最便捷路径。 AI治理框架与全球法规协同落地 (15:10) 应对董事会、外部审计与监管机构的要求，企业AI治理必须具备三项核心能力：大幅缩减非受控AI的受攻击面、量化并管理人类操作者与AI智能体的复合风险，以及在所有通信渠道中保持完整的数据溯源。一旦智能体发生越界或数据泄露，安全系统必须能追溯至向Microsoft Copilot或Anthropic Claude等提交的具体提示词、发起人以及事件演进链条。 随着合规要求由最佳实践转变为强制规范，Garruba建议企业将三大主流标准融为一体：以具备类似GDPR域外管辖效力的欧盟《人工智能法案》（EU AI Act）确立法律合规基准，采纳美国NIST AI风险管理框架（NIST AI RMF）指导日常运维方法论，并参照首个可认证的国际标准ISO/IEC 42001构建审计管理体系。Mimecast通过提供30天概念验证评估，协助企业建立AI用户图谱、定位外发数据通道并识别高危人员名单，以应对92%的普通员工同样被AI赋能提速所带来的风险扩散。 Garruba总结强调，企业无法通过延迟应对或机械封锁来消除AI风险，而必须在人机协同加速的环境中建立可审计的治理闭环；然而，面对未来数以千计自主智能体并发执行的场景，企业如何在保障业务创新的同时动态维持细粒度权限控制，仍是全行业有待持续解决的核心课题。

Detailed notes — zh

## Background 本视频由 Ai4 频道于 2026 年 8 月 17 日发布在 YouTube 平台，由 Mimecast 北美产品营销高级总监 Giulian Garruba 带来主题演讲。视频简介指出，演讲聚焦于企业职场中泛滥的影子 AI（Shadow AI）与自主智能体所引发的数据泄露隐患，旨在指导安全团队如何从盲区走向透明化，在不扼杀创新的前提下建立起高效的 AI 治理与数据防泄露框架。 ## TL;DR 演讲系统剖析了内部威胁从影子 IT、影子 AI 向自主智能体演进的历程，指出 8% 的关键高风险员工与以“机器速度”全天候运行的智能体正在加剧核心代码及机密数据的外泄风险。讲者强调简单粗暴的一刀切封禁只会适得其反，企业应当像胡佛水坝设置泄洪阀一样，实施弹性的自适应控制与行为引导，让合规路径成为最容易走通的路径。最终，结合欧盟《人工智能法案》（EU AI Act）、NIST AI RMF 与 ISO/IEC 42001 等前沿合规标准，演讲为企业落地“可见性、自适应控制与合规治理”提供了完整的三步走行动方案。 ## Atmosphere & style 演讲整体呈现出专业沉稳、循序渐进且极具警示意义的技术布道风格，讲者将生动的历史隐喻（如汽车安全带、水坝导流）与触目惊心的真实遥测数据紧密结合。现场互动调查中全场仅一人能提供完整审计记录的鲜明反差，直观揭示了行业普遍面临的治理盲区，营造出强烈的危机感与紧迫感。随着对智能体作为“终极内部人”威胁模型的拆解，分享迅速从风险揭示过渡到务实清晰的应对架构，给听众带来兼具前瞻性与落地可行性的认知重塑体验。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=4s) 内部威胁正从传统的影子 IT 演化为影子 AI 与自主智能体（Agentic AI），企业安全必须从被动响应转向主动自适应防御。 - [01:44](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=104s) 技术变革遵循“探索普及-风险发现-治理强制”三阶段演进模型，现代 AI 的风险爆发与治理周期远快于以往任何技术迭代。 - [03:26](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=206s) 影子 AI 在组织内高度泛滥，80% 员工及 90% 安全人员承认使用非合规工具，AI 治理策略若超过 90 天未更新即面临失效。 - [05:02](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=302s) 绝大多数企业内的自主智能体处于无防护状态，叠加“8% 的关键风险用户引发 80% 安全事件”的人类行为风险，成倍放大了潜在威胁。 - [07:14](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=434s) 研发人员对 AI 的高频使用导致数据泄露真实发生，上传至外部 AI 工具的内容中有高达 43% 为核心源代码。 - [08:05](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=485s) Meta 内部智能体越权暴露事件表明，智能体能在不突破传统防火墙的情况下，以“机器速度”自主行动导致严重数据越权。 - [09:12](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=552s) 智能体拥有合法访问权限并全天候运转，已成为打破人类生理限制、永不停歇的“终极内部人”。 - [11:08](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=668s) AI 治理必须始于端到端可见性，安全团队除了掌握在用工具与连接器权限外，更需洞察员工绕过安全策略背后的真实业务意图。 - [12:27](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=747s) 简单粗暴封禁 AI 只会加速影子 AI 泛滥，企业应建立基于风险连续光谱的自适应控制与行为微引导机制，“让安全的路径成为最容易走通的路径”。 - [15:10](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=910s) 结合 EU AI Act 法律底线、NIST AI RMF 落地方法论与 ISO/IEC 42001 认证体系，构建完整、可追溯的全球 AI 合规审计框架。 - [17:02](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=1022s) AI 让普通员工也能以“机器速度”操作，安全团队亟需通过画像分析与持续监控，精准识别高风险用户与智能体行为轨迹。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=4s) 影子 AI 与智能体演进：内部威胁的新形态 Mimecast 北美产品营销高级总监 Giulian Garruba 在开场中指出，人工智能已经从最初的新奇事物迅速演变为董事会层面的核心议题。本次演讲的主题直指核心——“从阴影到聚光灯”（From Shadow to Spotlight），旨在探讨影子 AI（Shadow AI）在组织内部的实际使用状况，以及安全团队如何为其照亮盲区。在接下来的分享中，演讲将重点帮助企业建立全面的可视性，并提供一套可直接落地的治理行动方案。 内部威胁（Insider Threat）是一个长期存在的严峻挑战。随着技术发展，这一概念已经从早期的“影子 IT”（Shadow IT）演变到“影子 AI”，如今更叠加了“智能体”（Agentic）的全新维度。拥抱 AI 的同时，内部风险被成倍放大。企业必须从原先被动的应对模式转变为主动防御思维，以便在安全事件发生前掌握全局动态。整个议题将围绕三大主线展开：深入剖析影子 AI 与智能体带来的根本性变革、构建切实可行的三步治理框架、以及通过可视性与主动自适应安全实现闭环防护。 ```mermaid flowchart LR accTitle: 内部威胁演进路径 subgraph S1["IT 与 AI 风险演进"] A["影子 IT<br/>(Shadow IT)"] --> B["影子 AI<br/>(Shadow AI)"] B --> C["智能体风险<br/>(Agentic AI)"] end subgraph S2["安全防御范式转型"] C -.-> D["被动响应模式<br/>(Reactive)"] D ==> E["主动自适应治理<br/>(Proactive & Adaptive)"] end ``` --- ### [01:44](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=104s) 科技变革的三阶段规律：从汽车安全带到 AI 治理 每一次重大技术突破在重塑人类社会与工作方式的同时，都遵循着相似的演进规律。任何技术变革通常都会经历三个鲜明阶段：首先是**探索阶段**（Exploration Phase），新技术从雏形走向大众并成为主流；紧接着是**风险发现阶段**（Risk Discovery）；最后则是**治理与强制执行阶段**（Governance & Enforcement）。 以汽车的发展为例：人类从马车时代跨入汽车普及时代，随后在实际使用中发现了全新的风险——马车发生碰撞时的破坏力有限，而汽车高速碰撞则会导致乘客被直接甩出车外造成惨重伤亡。面对这种新风险，人类社会引入了安全带这一治理机制，并进一步通过法律法规进行强制佩戴约束。值得注意的是，纵观历史，技术从出现到完成治理的时间周期正在急剧缩短。安全与合规团队所面临的技术迭代与风险爆发速度，远超以往任何时代。 ```mermaid flowchart TD accTitle: 技术变革与治理演进的三阶段模型 subgraph P1["第一阶段：探索与普及 (Exploration)"] A1["马车时代"] --> A2["汽车发明并成为主流"] B1["传统办公"] --> B2["生成式 AI 爆发普及"] end subgraph P2["第二阶段：风险发现 (Risk Discovery)"] A2 --> A3["高速碰撞造成人员甩出与严重伤亡"] B2 --> B3["影子 AI 滥用与敏感数据失控外泄"] end subgraph P3["第三阶段：治理与强制执行 (Governance)"] A3 --> A4["发明安全带并立法强制佩戴"] B3 --> B4["建立 AI 治理策略与自适应安全防护"] end ``` --- ### [03:26](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=206s) 影子 AI 的泛滥与数据外泄危机 生成式 AI 已经彻底摆脱了小众技术的标签，全面步入主流应用。行业数据显示，80% 的企业已经在使用生成式 AI；而对于声称“从未使用”的另外 20% 企业，真相往往是管理层“根本不知道组织内部正在发生什么”。从 2024 年到 2025 年，生成式 AI 的使用量呈现出惊人的 9 倍爆炸式增长。 然而，伴随 AI 广泛普及而来的并非完善的使用规范。高达 80% 的员工承认自己在工作中使用未经批准、未获授权的影子 AI 工具；令人担忧的是，在信息安全（InfoSec）相关岗位中，这一比例甚至飙升至约 90%。在这类未受管控的 AI 使用中，57% 的用户承认曾将敏感数据上传至公共 AI 工具中，包括系统密码、核心源代码及财务报表等。Giulian Garruba 强调，任何企业如果超过 90 天未重新审查其数据保护、内部风险或 AI 治理策略，该策略实际上就已经彻底过时。 - **AI 普及度**：80% 企业正在使用 GenAI，未使用的 20% 普遍存在内部盲区。 - **影子 AI 占比**：80% 普通员工使用非合规 AI 工具，安全岗位人员使用比例高达 90%。 - **数据泄露风险**：57% 用户主动向 AI 工具输入过密码、源码或财务等敏感信息。 - **策略失效周期**：AI 策略若 90 天内未更新，即无法应对当前安全态势。 --- ### [05:02](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=302s) 智能体风险倍增与“8% 关键人类风险” 自主智能体（Agentic AI）的兴起进一步放大了安全风险。IDC 预测，到 2029 年企业内部运行的智能体数量将突破 10 亿大关。智能体不再是传统的问答式聊天机器人，而是具备自主决策与执行能力的实体。企业招聘员工往往需要严格的背景审查和履历筛选，但在引入智能体时，人们通常只是直接下达指令：“这里有个问题，下周二前解决它，谢谢。”目前在财富 500 强企业中，仅有六分之一的智能体接受了安全防护，绝大多数智能体完全处于无防护状态运行。 在探讨 AI 风险时，绝不能脱离“人”这一核心。Mimecast 的运营经验表明，企业中通常 **8% 的用户导致了 80% 的安全事件**。随着市场整合、企业裁员以及“要求个人借助 AI 完成更多工作”的趋势加剧，这 8% 的核心风险人群被赋予了更高的系统权限和更大的工作职责。高风险群体往往包括掌握高权限的资深管理人员、接触核心财务数据的人员、即将离职并企图带走资产的员工，或是缺乏安全意识的粗心员工。因此，评估 AI 风险的真正突破口必须始于对人类行为的洞察。 --- ### [07:14](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=434s) 遥测数据揭示的源码与敏感数据外流 来自 Mimecast 内部风险防护平台的实时遥测数据显示，企业员工与外部 AI 工具之间的交互频率十分惊人。在受监控的环境中，平均每小时发生超过 5,500 次复制粘贴操作，产生 2,700 次文件上传，每月交互总量高达 290 万次。 最为触目惊心的数据是：上传至 AI 工具的内容中有 **43% 是源代码**。这一现象的根本原因在于研发人员和软件工程师始终是企业中最积极拥抱 AI 的群体。工程师为了以最快速度完成编码交付，频繁将代码库片段投喂给公共模型。这表明数据泄露并非潜在风险，而是每时每刻都在真实发生的现状；核心问题在于安全团队自身是否具备对此类行为的端到端可视性。 - **高频交互**：每小时 5,500+ 次剪贴板复制粘贴操作。 - **文件外发**：每小时 2,700+ 份文件直接上传至 AI 平台。 - **源码外泄**：AI 输入数据中 43% 为核心源代码。 - **月度规模**：受测企业群体每月与 AI 工具交互达 290 万次。 --- ### [08:05](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=485s) Meta 内部真实案例：机器速度下的权限越界 为了说明智能体风险的独特性，Giulian Garruba 分享了 2026 年初 Meta 发生的一起内部真实安全事件。一名内部员工在 Meta 的受限技术论坛上发布了一个提问，该内容原本仅供特定团队内部查看。然而，负责解答该问题的 AI 智能体错误理解了请求意图，自主将包含敏感技术细节的回答转发到了一个权限更宽泛的内部公开论坛，导致数千名原本无权访问的员工获得了直接访问权限。 尽管该内容在发布 2 小时后被发现并撤下，但 Meta 仍将其定性为最高级别的重大安全暴露事件（Critical Exposure）。该事件深刻揭示了智能体风险的本质：智能体并没有被外部黑客利用，也没有突破任何传统防火墙，更没有违背网络访问规则。它只是“以机器的速度自主行动”，并在执行过程中超出了预设边界。这种全新的威胁模型暴露了传统基于边界和规则的安全工具在智能体时代的无能为力。 ```mermaid flowchart LR accTitle: Meta 内部智能体数据越权暴露流程 A["员工在受限论坛<br/>提交技术提问"] --> B["AI 智能体<br/>自主处理与解答"] B -- "误解请求指令<br/>自主跨权限越界" --> C["转发至全员公开论坛<br/>(数千名无权员工可见)"] C --> D["定级为最高严重级别<br/>(Critical Exposure)"] ``` --- ### [09:12](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=552s) 智能体成为“终极内部人”与治理框架的开启 智能体正在演变成企业的“终极内部人”（The Ultimate Insiders）。正如惊悚电影中的经典台词所言——“电话是从房子内部打来的”（The call is coming from inside the house）。智能体天然拥有合法的系统访问权限，熟知企业内部架构、业务流程与组织架构树。与人类相比，智能体不眠不休、绝不犹豫、不知疲倦，且以指数级规模和机器速度并行运作。 在传统安全认知中，企业有一句行话：“大多数人类风险都发生在手指与键盘之间。”当员工深夜下班入睡后，由于没有人在邮件或聊天工具中点击恶意链接或下载文件，安全事件通常极少触发。然而智能体打破了这一生理限制，它们 24/7 全天候运转，使内部威胁升级为永不停歇的“机器速度风险”。针对这一严峻现实，演讲正式引入 AI 治理的核心三要素：**可视性（Visibility）、控制力（Control）与治理机制（Governance）**，并现场向参会的安全与合规负责人提问，展开对组织内部 AI 可接受使用策略（AUP）执行现状的探讨。 > “智能体将成为终极内部人。它们拥有对数据的合法访问权，熟悉你的系统和组织架构……智能体不会疲倦，不会怀疑自己的决定，它们正以机器的速度和指数级规模运行。” ### [11:08](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=668s) 治理第一步：从可见性（Visibility）开始与 CISO 的四个核心问题 讲者首先与现场观众进行互动调研，询问有多少企业认为自己已经具备完善的控制措施，能清楚掌握组织内正在使用哪些 AI 工具（无论是经过批准的还是未经批准的），以及这些工具应当如何规范使用。现场举手的人寥寥无几；随后讲者进一步提问，在这些举手的人当中，如果面临合规审计，有谁能够完整提供每天员工输入的所有 AI 提示词（Prompts）以及潜在的数据外泄（Exfiltration）详细记录？台下最终只剩下一只手举着。这生动地证明了绝大多数企业在面对生成式 AI 时，在安全和可见性层面依然有很长的路要走。 为了打破这种盲区，讲者提出任何 AI 治理策略都必须始于“可见性”。每位首席信息安全官（CISO）都应该能够回答以下四个关键问题： 1. **组织内正在使用哪些生成式 AI 工具？** 2. **当前有哪些 AI Agent（智能体）在运行？具体是谁在运行它们？** 3. **AI 连接器（Connectors）拥有什么权限？**（包括像 MCP 服务器、通过企业采购的第三方商业软件所集成的智能体等） 4. **你是否理解员工使用这些工具背后的真实意图（Intent/Why）？** 讲者特别强调了第四点的重要性：如果员工选择绕过企业的安全政策去使用影子 AI（Shadow AI），根本原因往往是企业现有的官方工具无法满足其实际工作需求。例如，当员工尝试用官方指定的 AI 模型生成某种图像或视觉素材却始终达不到预期效果时，他们自然会跳去使用其他未受监管的外部模型。如果企业只提供单一死板的选择，实际上是在为内部人风险（Insider Risk）和未经授权的 AI 滥用大开方便之门。 ```mermaid flowchart TD accTitle: AI治理与自适应控制流程图 accDescription: 展示从可见性发现、自适应控制判定到教育引导与合规审计的完整治理路径。 subgraph Visibility["1. 可见性与意图洞察 (Visibility)"] V1["识别在用生成式 AI 工具"] V2["监控运行中的 AI Agents 与执行者"] V3["审查 MCP 与第三方连接器权限"] V4["分析员工真实业务意图"] end subgraph Control["2. 自适应风险控制 (Adaptive Control)"] RiskSpectrum{"风险等级评估"} LowRisk["低风险: 公开营销资料等<br/>-> 放行并记录"] MedRisk["中风险: 违规外部工具<br/>-> 行为推送引导 (Nudge)"] HighRisk["高风险: 核心代码/未公开财报<br/>-> 源头/目标拦截与隔离"] end subgraph Governance["3. 审计与合规治理 (Governance)"] G1["EU AI Act 法律合规"] G2["NIST AI RMF 风险方法论"] G3["ISO/IEC 42001 体系认证"] end Visibility --> RiskSpectrum RiskSpectrum --> LowRisk RiskSpectrum --> MedRisk RiskSpectrum --> HighRisk Control --> Governance ``` --- ### [12:27](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=747s) 自适应控制与行为引导：不要把 AI 当成非黑即白的开关 在建立可见性之后，治理的下一步是进入“控制（Control）”阶段。讲者指出，企业在管理 AI 时可能犯的最大错误就是将其视为一个非开即关的“通断开关”（On-Off Switch）。简单粗暴地全面封禁 AI 不可能解决问题，正确的做法是采取“自适应方法（Adaptive Approach）”，因为安全风险本身就是一个从无风险、低风险、中风险到高风险的连续光谱（Risk Spectrum），安全控制措施也必须具备相对应的弹性。 例如，市场部门引用公开统计数据让 AI 辅助撰写一篇博客文章，其风险微乎其微；而如果一个自主运行的 AI Agent 擅自访问组织内部尚未发布的季度财报数据，则属于极高危事件。自适应控制的实施主要体现在以下几个维度： - **按数据源拦截（Block by Source）：** 锁定组织内绝对不可泄露给 AI 的核心资产（如私有代码库、核心财务系统），直接切断这些系统向外传输的通道。 - **按目标拦截（Block by Destination）：** 明确区分受认可的安全 AI 平台与明令禁止访问的不可信外部工具。 - **极端风险处置：** 当行为严重越轨时，采取终端隔离（Quarantining Endpoints）或直接剥夺系统访问权限。 此外，AI 治理绝不仅是惩罚与阻断，更是“教育与引导（Education）”。Mimecast 提供了诸如行为微提示（Behavioral Nudges）等机制，例如当员工使用了未经批准的 AI 工具时，系统会立即在 Slack 上发送一条温和的提醒消息予以指引。更具前瞻性的是，这种教育对象在未来几个月内将不仅限于人类员工，还将涵盖 AI Agent 本身——因为智能体往往不会完全遵循既定护栏，企业必须在约束它们的同时对其进行训练和规范。 讲者引用了 Google Cloud CISO 办公室的名言：“一味封禁生成式 AI，而不是为组织安全使用它开辟一条清晰明亮的道路，往往是徒劳无功的，这甚至会加速影子 AI 的泛滥和数据泄露，反而招致企业原本极力避免的风险。”讲者以胡佛水坝（Hoover Dam）为例进行了生动的类比：小孩子在溪流或海滩筑起沙墙挡水，水流总会从四周绕过去；而真正的大坝绝非死堵，它们在山体一侧设计了泄洪减压阀（Relief Valves），从而实现对水流的主动引导与控制——“让安全的路径成为最容易走通的路径（Make the safe path the easy path）”。 --- ### [15:10](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=910s) 全球合规框架与审计需求：EU AI Act、NIST 与 ISO 42001 当企业向董事会、审计机构或监管部门汇报时，AI 治理框架必须能够清晰证明三件事： 1. **降低企业内未经授权 AI（Unsanctioned AI）的风险；** 2. **准确洞察人类员工与 AI Agent 所带来的双重风险；** 3. **无论通过何种渠道，都能保护并审计数据输出。** 讲者强调，一旦发生安全事件（例如某人启动了一个 AI Agent，而该智能体突破了边界导致数据外泄），安全团队必须能够精确调取当时的原始记录，指出“正是这段发送给 Copilot 或 Claude 的对话生成了该 Agent，其操作者是谁”。 这些要求在当前显得尤为紧迫，因为就在本次演讲的两天前，**欧盟《人工智能法案》（EU AI Act）**已正式生效。该法案与 GDPR 类似，其管辖范围不仅局限于欧盟本土企业，更涵盖所有向欧盟市场投放 AI 产品的全球实体。面对复杂的全球监管环境，讲者梳理了三大核心框架的定位： - **EU AI Act（欧盟人工智能法案）：** 必须严格遵守的强制性法律底线。 - **NIST AI RMF（美国国家标准与技术研究院 AI 风险管理框架）：** 虽然属于自愿性质，但正迅速被美国政府机构及各行业采纳为事实上的标准方法论。 - **ISO/IEC 42001：** 全球首个针对人工智能管理体系（AIMS）的可认证国际标准。讲者提到，Mimecast 是首家获得该标准认证的网络安全公司。 讲者建议企业将三者融合考量：以 EU AI Act 明确法律底线，以 NIST 框架作为落地方法论，以 ISO 42001 构建可供第三方审计的规范化体系。 --- ### [17:02](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=1022s) 人人提速时代的内部人风险与 30 天验证方案 在演讲的尾声，讲者重新将焦点拉回到“人”的因素上。企业安全领域有一个经典的二八法则：**8% 的高风险用户往往制造了组织内 80% 的安全风险**。在 AI 时代，这 8% 的人正在以前所未有的速度借助新技术狂奔；但更严峻的挑战在于，其余 92% 的普通员工也同样在以前所未有的速度行动，他们手握更多工具，并以“机器的速度（Machine Speed）”进行操作。这意味着企业突然之间要面对成千上万个具备超强行动力的新型“内部人”。 为了帮助企业摸清底数，讲者介绍了 Mimecast 提供的 30 天无风险价值验证（Proof of Value）方案。该方案能够协助企业全面看清： - 组织内哪些具体人员正在使用哪些数据和 AI 工具； - 数据正在被共享发送至哪些外部目标端； - 运行中的所有 AI Agents 及其具体行为轨迹。 通过按姓名、团队和地理区域清晰绘制出那 8% 核心风险人群的用户画像，安全团队才能真正掌握由 AI 引发的内部人风险。最后，讲者在全场掌声中结束了演讲，并预留了问答时间。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=4s) 从阴影到聚光灯（From Shadow to Spotlight） — Giulian Garruba > [03:26](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=206s) 任何企业如果超过 90 天未重新审查其数据保护、内部风险或 AI 治理策略，该策略实际上就已经彻底过时。 — Giulian Garruba > [05:02](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=302s) 这里有个问题，下周二前解决它，谢谢。 — Giulian Garruba > [05:02](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=302s) 8% 的用户导致了 80% 的安全事件 — Giulian Garruba > [09:12](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=552s) 智能体将成为终极内部人。它们拥有对数据的合法访问权，熟悉你的系统和组织架构……智能体不会疲倦，不会怀疑自己的决定，它们正以机器的速度和指数级规模运行。 — Giulian Garruba > [09:12](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=552s) 电话是从房子内部打来的（The call is coming from inside the house） — Giulian Garruba > [09:12](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=552s) 大多数人类风险都发生在手指与键盘之间。 — Giulian Garruba > [12:27](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=747s) 一味封禁生成式 AI，而不是为组织安全使用它开辟一条清晰明亮的道路，往往是徒劳无功的，这甚至会加速影子 AI 的泛滥和数据泄露，反而招致企业原本极力避免的风险。 — Google Cloud CISO 办公室 > [12:27](https://www.youtube.com/watch?v=BWtXPtvRlW8&t=747s) 让安全的路径成为最容易走通的路径（Make the safe path the easy path） — Giulian Garruba ## Mentioned Giulian Garruba, Mimecast, Meta, Google Cloud, IDC, EU AI Act, NIST AI RMF, ISO/IEC 42001, GDPR, MCP, Slack, Copilot, Claude ## Video description Giulian Garruba - Senior Director, Product Marketing, Mimecast North America AI is everywhere in the workplace, and so is Shadow AI. Unsanctioned tools and autonomous agents are quietly opening new paths for data exposure. This session covers moving from blind spots to visibility: understanding how employees interact with GenAI and AI agents, building governance frameworks that don't kill innovation, and preventing data leaks before they hit headlines. Get actionable guidance on how to accelerate productivity & innovation with AI without losing control of your data. [Source](https://www.youtube.com/watch?v=BWtXPtvRlW8)

Source quotes

[0:04] Giulian Garruba: 从阴影到聚光灯（From Shadow to Spotlight） [3:26] Giulian Garruba: 任何企业如果超过 90 天未重新审查其数据保护、内部风险或 AI 治理策略，该策略实际上就已经彻底过时。 [5:02] Giulian Garruba: 这里有个问题，下周二前解决它，谢谢。 [5:02] Giulian Garruba: 8% 的用户导致了 80% 的安全事件 [9:12] Giulian Garruba: 智能体将成为终极内部人。它们拥有对数据的合法访问权，熟悉你的系统和组织架构……智能体不会疲倦，不会怀疑自己的决定，它们正以机器的速度和指数级规模运行。 [9:12] Giulian Garruba: 电话是从房子内部打来的（The call is coming from inside the house） [9:12] Giulian Garruba: 大多数人类风险都发生在手指与键盘之间。 [12:27] Google Cloud CISO 办公室: 一味封禁生成式 AI，而不是为组织安全使用它开辟一条清晰明亮的道路，往往是徒劳无功的，这甚至会加速影子 AI 的泛滥和数据泄露，反而招致企业原本极力避免的风险。 [12:27] Giulian Garruba: 让安全的路径成为最容易走通的路径（Make the safe path the easy path）

Source quotes

[0:04] Giulian Garruba: From Shadow to Spotlight [5:03] Giulian Garruba: Here's a problem. Solve it by Tuesday. Thank you [5:03] Giulian Garruba: do more with AI [9:12] Giulian Garruba: The call is coming from inside the house. [9:12] Giulian Garruba: between the fingers and the keyboard [12:27] Giulian Garruba: block by source [12:27] Giulian Garruba: block by destination [12:27] Google Cloud Office of the CISO: blocking generative AI rather than creating a well-lit path for organizations to use it and securely is ineffective. It often precipitates the proliferation of shadow and data leakage—the very risk enterprises were trying to avoid. [12:27] Giulian Garruba: the safe path is the easy path.

Original transcript

[0:04] As he mentioned, my name is Julian Geruba.  I lead the product marketing team at MCAST. [0:09] And I'm excited to talk to you today about uh  the explosion we've seen obviously uh of AI [0:15] going from a curiosity all the way  to a board level conversation um and [0:20] what that means for your organization. Uh  the title is very deliberate. Um, we are [0:26] really going to talk about AI in the shadows  here, how it's being used in your organization, [0:31] as well as how we can shine a light on it. So,  for the next 20 minutes, uh, we'll give you some [0:36] ways to get visibility, and then we'll  talk about a plan you can take back to [0:39] your organizations to ensure  you're seeing what's going on. [0:44] So, we're going to talk about the pressing  challenge of insider threats. Now, this is a [0:49] concept that's been around uh for a long time. um  it's evolved from shadow IT to shadow AI and now [0:56] it's taken on an agentic spin. And so uh as we  embrace AI, the risk of insider threat becomes [1:03] larger. And so we're going to talk talk today  about a shift from reactive to proactive thinking [1:09] so you can get ahead of this insider threat  and have visibility into what's going on. [1:14] So we're going to do that in three ways. The  first we're going to explore the fundamental [1:18] change uh of Shadow Aonogentic. Second, we're  going to talk about a government governance [1:23] strategy, excuse me. Um, a framework that you can  think about and and three steps you can take back [1:28] to your organizations. And then finally, we're  going to talk about solving this problem uh by [1:33] getting visibility and proactive adaptive security  on top of that. Can just quickly show a hand. How [1:39] many of you in the room are security or  compliance leaders in your organization? [1:44] Awesome. Cool. So starting off, let's start  with the risk side of AI because before we talk [1:50] about governance, it's really important that we  understand clearly what we're actually governing. [1:56] Now let's talk about transformations and  technological advances we've seen in the world. [2:02] Um, every one of these eras has transformed  everything. You know, the way we work, [2:08] the way we go through society, but  it's also transformed the way we work. [2:13] And I apologize for the the misformatting on the  the template, but each of these transformations [2:19] have has had three distinct phases. And  I'll give you the car as a good example. [2:23] The exploration phase where we went from horse  carriages all the way to starting to see cars [2:29] on the road to them becoming mainstream. Everyone  has one. Somewhere along that line, we had a risk [2:34] discovery. What was that risk discovery? Well,  it turns out when a horse and buggy crashes, [2:40] uh, there's a lot less damage done than when a car  crashes and people are flung out of the vehicle. [2:45] So, what was the risk we discovered and what was  the governance that we put in place? We decided [2:50] we needed seat belts, right? So, we needed a  seat belt. Uh, and we didn't just put a seat [2:55] belt in every car. We also made sure that you get  enforcement on that. So, then sure that it's not [3:00] um, you know, you're not wearing you're not not  wearing it in your car. Um, but look at the bottom [3:06] and notice the time scale that's changed. Um, and  how fast these evolutions have gotten smaller. [3:13] Security and comp compliance teams are always  evolving and they've been this way for decades, [3:17] but these things are happening faster and faster  than ever before. And I know all of you guys [3:22] understand this and know this  because you live it every day. [3:26] So, generative AI absolutely no longer a  niche. It's become completely mainstream. [3:32] Um, and what we see is that 80% of organizations  are using generative AI, but what we know [3:38] is that the 20% that say they aren't truly just  don't know what's going on in the organization. [3:44] And a nine times explosion of general AI use from  2024 to 2025. And if you continued that forward, [3:51] you'd see a larger number as well. But if the  use of AI is growing, then surely all of that [3:57] is being done within well- definfined policies,  right? Well, not so much because we see that 80% [4:03] of users admit to using shadow AI, meaning using  tools that are not approved, that they don't have [4:10] permissions to use, or that the organization does  not allow. The sad part is that in infosc roles, [4:15] that's about 90%. 57% of those users admit to  sharing sensitive data into those AI tools. [4:26] And the reason that I think this one is  interesting is if you think about the types of [4:30] users in your organization uh that are using the  sensitive data, these are just the ones that are [4:36] admitting things like passwords, we all get it.  Uh source code, financials, all being shared. So [4:42] if you have a data protection, an insider risk  or an AI strategy, it's probably outdated if you [4:46] haven't res reviewed it in the last 90 days and  it will continue to be something that you need [4:50] to build a policy around evolving and continuing  to move faster and faster. Now, we talked about [4:57] shadow AI. That's kind of the human side of it,  but Agentic absolutely multiplies that risk. [5:03] IDC projects that by 2029, there will be over a  billion agents uh in organizations. And the thing [5:10] I like to think about this is it's not chat. These  are not just chat bots. These are not things that [5:15] you're giving a call in response like a generative  AI tool. These agents are decision makers. [5:20] the way that we would hire an employee and do a  background check and look for a certain resume. [5:25] We're not doing that with agents. We're just  saying, "Here's a problem. Solve it by Tuesday. [5:29] Thank you." One in six of those agents is secured  in the Fortune 500, which means on the flip side, [5:35] the vast majority are not secured. And when  you think about forecasting this forward today, [5:41] each user has a couple of agents that they're  using. That quickly becomes thousands in the next [5:46] couple months. Now, MCAST is a securing human data  and AI company. And so, we already take a lot of [5:53] information um from our learnings with customers  as we help them get a handle on their human risk [5:59] strategy. And what we've seen is that 8% of users  typically generate 80% 80% of security incidents. [6:08] Now, if you think about what we're seeing  in the market with the consolidation, [6:11] we're seeing layoffs, we're seeing people  being asked to do more and more with AI, [6:16] that 8% is changing because the individual user  is taking on the role of more and more people. [6:25] That means they have more impact.  They have more responsibility. [6:28] And if we're not training them appropriately, if  we're not focused on what that human is doing, [6:33] we're missing a big part of of the risk. So  the point here is that risk is not uniform. [6:39] The people that typically cause that 8%  are people with high access high levels of [6:43] access to your organization. They may be in  finan senior roles. They may be in you know [6:48] financial roles things access to sensitive  data. Um they may be departing employees [6:54] who know they're leaving and are looking to  take something with them. Or they may just [6:58] be careless and not really understand what the  problem that they're is that they're causing. [7:02] Now this may seem counterintuitive but  my suggestion is that in order to get a [7:06] handle on this in your organization to get a  true picture of AI risk start with the human. [7:14] So from our AI tools and our insider risk  platform what we see is that every hour over 5,500 [7:23] that's so small I can't even see it. Uh so my  apologies there. 5,500 copy and paste per hour. [7:29] um 2,700 files uploaded per hour. And the  one that scares me the most is this 43% [7:34] of what's going into those tools being  source code. Think about it. It makes sense. [7:39] The people typically leading the charge in your  organization tend to be developers, engineers. [7:44] And what do they have access to? What are they  trying to get done as fast as possible? Coding. [7:49] And so that one is really concerning. Overall,  it comes back to 2.9 interactions per month. [7:55] 2.9 million, excuse me. So the question is  not is this happening in your organization, [8:01] it's whether or not you have visibility into it. [8:05] So we talked about humans, we talked about shadow  AI. What about those agents? Now I'll show you a [8:09] real example uh from earlier this year that  helps you put the agentic risk into focus. [8:14] Uh so everyone here is familiar with Meta. Early  in 2026, they had an incident where an employee [8:20] posted a question on an internal technical forum,  not extend intended to be exposed externally [8:27] and it wasn't. But what happened is  the agent that answered that question [8:32] misinterpreted the request and post it  on an even more public forum internally [8:37] where thousands of employees that shouldn't have  had access had access. and it was only there for [8:41] 2 hours, but Metaclassified it as the highest  level of critical exposure that they have. [8:48] Why this matters is because the agent didn't  break any rules. It didn't get through a [8:54] firewall. It didn't come from the outside. It  just simply acted autonomously at machine speed. [9:01] It operated outside of its intended boundaries.  And this is what we should expect is the new [9:06] threat model. And traditional security compliance  tools are just simply not built for it. [9:12] So this is the key thing I want  you to take away from this. [9:15] Agents will become the ultimate insiders. Think  about it. They have legitimate access to your [9:21] data. They know your systems. They know your org  chart. They're inside your perimeter. The call is [9:27] coming from inside the house. Now, why is this a  problem? Because it's happening at machine speed. [9:35] And we talked about the the things that our humans  are doing. Agents don't get tired. They don't [9:40] second guessess their own decisions. They operate  at exponential scale. So think about the challenge [9:45] that we face when we think about this risk uh  landscape. This is insider risk at machine speed. [9:52] So I talked about the consolidation with you  have fewer employees with more permissions. [9:56] The problem is not you have fewer insiders. The  problem is that you have many more insiders. [10:02] And those agents are users as we talked about  with a set of permissions that you are setting [10:06] with behavior to monitor. But they're not  bound by those traditional limitations. [10:11] We have a a phrase we like to say at at Mcast that  the the uh between the fingers and the keyboard is [10:17] where most of the human risk lies because when  your organization is at home asleep not working, [10:23] most security incidents are not being kicked  off during that. They might in the morning [10:27] you come on, you come in, you open up a team's  message or an email, you click on something, [10:31] you download something, you share something. But  agents don't have this problem. They're working [10:37] 24/7 around the clock. So, we spent enough time  talking about the risk. What do we do about it? [10:45] So, let's talk about governance. And really, this  boils down for me to three things. Visibility, [10:50] control, and governance. Now, I did a show of  hands earlier on security and compliance leaders. [10:57] How many of you think with a show of hands that  your organization has an AI acceptable use policy [11:03] in effect today? Okay, keep your  hand hands raised. Of those that do, [11:08] how many think they have controls in place  to understand which AI tools are being used, [11:13] whether they're sanctioned or unsanctioned, and  how they should be used? Couple still up. Finally, [11:18] how many of you think of those that your  organization could provide details during an audit [11:24] of all the excfiltration or all  the AI chats that are being uh [11:29] entered every single day? I think I got one  still up. So we have a long ways to go on this. [11:36] My suggestion is we start with visibility. There  are four cush questions up here that every CISO [11:41] should be able to answer. The first which  generative AI tools are in use in my organization. [11:48] The second which agents are running who's running  them. The third is what access do AI connectors [11:54] have? I mean things like MCP servers, commercial  agents delivered through thirdparty products you [11:59] consume. And finally, do you understand the why,  the intent behind the tools? Because if users are [12:07] circumventing your policy, it's probably because  they don't have what they need. All of us have [12:11] been through a situation where you've tried to use  an AI model. You're not getting the graphic or the [12:16] visual you want. You jump over to a different  one. If you only have one that you're offering, [12:21] you're opening yourself up to instant insider  risk, unsanctioned AI use in your organization. [12:27] So, start with visibility. And then we move  into control. The biggest mistake you can make, [12:32] and I think all of us in this room understand  this, is treating AI as an on-off switch. We [12:36] would never just take it away because that's not  going to solve the problem. What we need to do [12:41] is take an adaptive approach because risk lies on  a spectrum from no risk to low to medium to high. [12:48] And your controls should do the same thing. Think  about a blog that is shared by the marketing [12:52] department that's referencing, you know, publicly  available statistics. significantly less risk [12:58] than an agent autonomously accessing unreleased  financial data for the quarterly earnings report. [13:05] Those are big differences and adaptive controls  will help you manage this differently. So when we [13:09] think about things like block by source, meaning  if I can identify in my organization areas of data [13:15] that I think should never make it out to AI, maybe  it's a code base, financial data, I can block [13:20] those systems. Second, I can look at destination.  So maybe there's certain AI tools that I know AROC [13:26] I'm okay with and others that I absolutely know  we do not use. So if I can establish both block by [13:31] source and block by destination, then all I have  to do is really address when things really get [13:37] off the rails. Things like quarantining endpoints,  removing system access when it's really important. [13:42] But it's also not just about control. It's also  about education. And and at MCAST, we have a lot [13:47] of of of to of tools to help do that. some things  like behavioral nudges that send a Slack message [13:52] to say, "Hey, you just used an enterprise an  unimproved AI tool, but it's also becoming a [13:57] conversation about education, not just for that  user, but also for the agent. So something to [14:02] think about in the next couple months as we start  thinking about what are the ways that we can guide [14:06] an agent. Yes, they don't follow the guard rails  all the time, but if we can be restricting them [14:12] and actually training them in addition to  training the user that's operating them, uh, [14:16] we can do a lot." Thank you. Here's a quote I like  from the Google Cloud Office of the CISO. Blocking [14:22] generative AI rather than creating a well-lit  path for organizations to use it and secur and sec [14:28] and securely is ineffective. It often precipitates  the proliferation of shadow and data leakage. [14:34] The very risk enterprises were trying to avoid.  Remember when you were a kid and you were in a [14:39] playing in a stream or a creek or the beach and  and the water was running and you'd put your hand [14:44] in or you'd build a wall and what happens to that  water? It finds a way. It goes around. Think about [14:50] the real dams. There's a very famous dam about 30  miles from here. They don't take that approach. [14:55] They have relief valves built in to the to the  side of the mountain that to that dam so that [15:00] they can control the way that that information  is going out. Make the safe path the easy path. [15:10] So, finally, governance. You need to be able  to show a board, an auditor, a regulator [15:16] that you can do three things. One, cut the risk  of unsanctioned AI in your organization. Two, [15:22] understand the risk of the human and the  AI agents. And three, protect and produce [15:29] data regardless of the channel. Because if a  security incident happens and it's a result of [15:35] someone kicking off an agent and that agent not  being able to follow boundaries or guardrails and [15:40] data gets excfiltrated, being able to say here is  the exact chat that was sent to Copilot or Claude [15:47] that created that agent and who it was.  That's that's all becoming very critical. [15:52] And I'll show you why. Because this is actually a  very timely day uh to be giving this presentation [15:57] because two days ago the EU AI act went into  effect. And I know there's a laundry list of words [16:03] up there. Essentially, this doesn't just mean  that I have to be careful about the way I'm using [16:10] AI if I'm an a EU based company. It also applies  to anyone placing AI on the EU market. It's very [16:17] similar to GDPR. If you're doing business in  these regions, you need in these regions, it [16:21] becomes something you need to care about. The one  in the middle is a US standard. It's a voluntary [16:26] framework that NIST put out. Um, but it's being  adopted by government agencies pretty quickly as [16:31] the deacto standard uh for AI. And finally, ISO  4201 is the first certifiable standard for AI. [16:40] And Minecast was actually the first cyber security  company to achieve this. Um, my recommendation is [16:45] think about all of these things together. They  may not all apply to you, but if you think about [16:49] the EU AI act as those legal requirements  that I have to comply with, think about the [16:54] NIST framework as a methodology you can follow  and the ISO framework uh as an auditable system. [17:02] Coming back to that human piece, remember that  the 8% of users create 80% of risk in your [17:09] organization. That 8% is now moving faster than  ever. But here's the problem. The 92% are also [17:15] moving faster than ever. They have more tools  and they can operate at machine speed. So the [17:21] question becomes, are you ready for thousands of  new insiders in your organization? Now, not here [17:27] to do a product pitch, but if you're curious, uh,  Mcast does have a 30-day risk-free proof of value [17:33] that will help you understand the users that  are using data and AI in your organization, [17:38] the destinations, the destinations, all of that  data is being shared to, all of the agents, [17:42] and what they're doing. um so that you can get a  clear picture of that 8% in your organization by [17:48] name, by team, by geography, and really start to  get a handle on the insider risk caused by AI. [17:55] So, we're out of time. I have about two minutes  left. Um I can take one or two questions, [17:59] and then I'm happy to take more after time ends  out out in the hall. But thank you, [applause]