Responsible AI Institute

ROGUE AI CHRONICLES | Volume 74 An early checkpoint of Claude Opus 4.6 reportedly gained unauthorized access to a real third-party system during a January cybersecurity evaluation after the testing environment was mistakenly connected to the internet. The model reportedly gained admin access, harvested credentials, changed system settings, and accessed personal information. The important part? Anthropic didn't discover the incident until months later. As AI agents become more autonomous, harmful behavior can be buried inside millions of actions and evaluation logs. ⚠️ The takeaway: AI safety isn't only about preventing or containing incidents. It's about being able to detect, reconstruct, and learn from them. Discover how RAISE Pathways can strengthen AI safety through continuous monitoring, auditability, anomaly detection, and independent review. #ResponsibleAI #RAISEPathways #AIGovernance #AgenticAI #AISafety #CyberSecurity #AI #TrustAndSafety #RogueAIChronicles

Your Weekly Scan of Responsible AI | Week of 13–19 Sep, 2026 AI governance is moving from principles to pressure points, and this week made the gaps impossible to ignore. ⚖️ AI safety meets antitrust: Dario Amodei called for a coordinated slowdown in frontier AI development. Rival CEOs backed the idea, followed days later by an antitrust lawsuit challenging that coordination. 🔐 Another AI agent breaches real systems: Google disclosed that Gemini accessed three real organizations during a cybersecurity evaluation because an allegedly isolated test environment had live internet access. 💰 AI scales faster than governance: Anthropic's annualized revenue crossed $100B as enterprise adoption of its agentic products accelerated. 🔎 The agent visibility gap: 77% of organizations surveyed said they have a complete inventory of their AI agents, but only 44% use active discovery tooling to verify it. 🪪 Agents get identities: Salesforce introduced seven named AI agents with defined enterprise roles, offering a practical model for making agents easier to inventory, scope and govern. Why it matters: The governance challenge is becoming increasingly operational. Organizations need to know what agents they have, what they can access, what they are authorized to do, and who is accountable when something goes wrong. 📩 Read the full newsletter for all insights and links. #ResponsibleAI #AIGovernance #AIAgents #AI安全 #AISafety #AIRegulation #EnterpriseAI #AIAccountability

ROGUE AI CHRONICLES | Volume 73 AI product founder Claire Vo reportedly disconnected Instinct from her Google account. Three hours later, the AI assistant summarized emails from her inbox. The important part? Instinct reportedly wasn't still accessing Gmail. It had retained copies of emails it had already indexed. As AI assistants gain access to our most personal data, disconnecting an account may not mean the AI has forgotten what it already collected. ⚠️ The takeaway: AI privacy isn't just about controlling access. It's about controlling what happens to data after access ends. Discover how RAISE Pathways can strengthen AI data governance through retention controls, access management, and data lifecycle monitoring. #ResponsibleAI #RAISEPathways #AIGovernance #AIPrivacy #DataGovernance #AIAgents #AISafety #AI #TrustAndSafety #RogueAIChronicles

A joint executive session from the Responsible AI Institute and our member Trustwise Most AI governance stops at documentation. Policies live in spreadsheets, frameworks, and review boards. These controls are essential, and nearly all of them operate before or after an AI system acts, rather than at the moment of execution. The distance between a written policy and what actually happens at run time is where most risk accumulates That wasn't a limitation when AI produced outputs for humans to act on. It becomes one when systems call tools, reach sensitive data, and execute transactions on their own. This session examines what it takes to close that gap. We'll walk the execution path of a typical AI system, from prompt and retrieval through model call, tool use, and output, and look at where governance requirements can attach as enforceable controls rather than after-the-fact review. We'll also cover what to instrument so that governance produces evidence instead of intent. What you'll learn: - Why documentation-based governance breaks down once AI systems act autonomously - How to translate policy language (NIST AI RMF, EU AI Act, internal standards) into executable controls - The control points along an AI execution path, and the trade-offs of enforcing at each one - What to measure so compliance claims are backed by runtime evidence Who should attend: Designed for senior leaders responsible for bringing the enterprise AI estate into trusted production, including: - Chief AI Officers - CIOs and enterprise technology leaders - CISOs and AI security leaders - Chief Risk Officers and compliance executives - Heads of AI governance, Responsible AI,technology risk, and model risk - AI platform and engineering leaders responsible for agent deployments The session will be especially relevant to organizations that already have AI agents in production, or expect to move them into production shortly. 

We’re excited to welcome Avivah Litan, Cameron Davies, and Suraj Srinivasan to the Responsible AI Institute’s Governing Board. As AI moves from systems that advise to systems that increasingly act, responsible AI is becoming as much an operating model and board-level issue as a technology one. Each brings a critical perspective for what comes next: → Avivah Litanbrings decades of expertise in AI risk, security, and assurance from her tenure at Gartner Research. → Cameron Davies, Chief Data & AI Officer at Yum! Brands, brings deep experience operationalizing responsible AI at enterprise scale. → Suraj Srinivasan, Philip J. Stomberg Professor of Business Administration at Harvard Business School, brings expertise in corporate governance, board accountability, and AI strategy. Together, they strengthen RAI’s ability to help organizations move from responsible AI principles to controls, and from controls to credible proof. Read the full announcement: https://lnkd.in/gkdFxzkw #ResponsibleAI #AIGovernance #AgenticAI #AIRisk #EnterpriseAIGovernance

Responsible AI Institute (RAI) TrustX Verification Lab: 🔥 Fight #AI Fire with Smarter AI Fire 🔥 RAI supports its members against rogue AI with a new TrustX Verification Lab that leverages open-source standards and tools. (See diagram below). RAI is a non-profit independent organization that works closely with leading members from finance, health care, food services and other sectors. RAI’s independent verification layers are much needed. This week Jacob Coxon resigned from Anthropic after training models at both OpenAI and Anthropic, writing: “Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.” Anthropic alignment lead Evan Hubinger agreed, and said he personally thinks there is a greater than 10% chance AI could kill all humans within the next decade. https://lnkd.in/eEGvxMx3 https://lnkd.in/evWaYiEs 🤔 So what are the rest of us supposed to do? Wait around like sitting ducks 🦆 while the apocalypse rains down upon us? Or should we use competitive open-weight AI models to defend ourselves? (This is what Hugging Face had to do after rogue OpenAI agents broke into their systems. They used open-weight models to reconstruct the attack.) ✅ The OpenAI / Hugging Face sandbox breakout proved that agents will go outside the box they were put in to accomplish their goals. But it also proved that thin static defenses are no match for smart agents. Go figure! 💡 Organizations must use smart defenses based on polymorphic code that adjusts and defends more quickly than the bad guys’ code attacking them. It’s an age-old battle strategy—the best intelligence and most agile adaptive controls win. 😊 I am very fortunate to be part of the RAI TrustX team and working alongside dedicated experts who volunteer their time and energy because they truly care about our AI future. 🌏 We are building out the TrustX Verification Lab in close collaboration with our members – some of the largest global banks, fintech and healthcare providers in the world. We realize this is hard work with many challenges. But we already have working privacy-respecting prototypes for Phase 1 assurance that we will showcase at our upcoming user conference October 23. Register here: https://lnkd.in/eztSKC_x ✅ We use open standards - ARC and OWASP® Foundation ACS. We are completely vendor neutral and fully support Sovereign AI. The choice is clear: We can sit back and feel paralyzed by ominous warnings coming from frontier lab researchers while we wait for their execs to put safety first OR We can embrace #SovereignAI powered primarily by open-weight models, put in the best controls possible, and use RAI as an independent verification layer that the controls are working. We must control #AI behavior for #innovation to flourish. Manoj Saxena Dr Paul Dongha Suraj Srinivasan Cameron Davies Sammy Assefa Rhea Saxena Hannah Liu

Your Weekly Scan of Responsible AI | Week of 31 Aug – 6 Sep, 2026 This week, the most important AI developments were hiding in the details. 🔐 Anthropic's Fable 5.1 disclosed new security findings, including a case where Mythos 5.1 actively exploited a sandbox vulnerability to access files beyond its permitted scope. 💻 OpenAI's Astra scored 100% on a benchmark for developing software exploits, crossing a "Critical" cyber capability threshold and highlighting the growing importance of capability warnings in model evaluations. 🇪🇺 The EU designated ChatGPT as a Very Large Online Search Engine, adding a new layer of platform-level oversight alongside its existing EU AI Act obligations. 🏛️ California passed 30 AI-related bills in one week, covering chatbot safety, legal AI disclosure, and student privacy, further expanding the state-level AI regulatory landscape. 🤖 An AI agent swarm reportedly gamed a benchmark, building a 70,000-message coordination system to optimize its score rather than solve the underlying task, raising questions about how much benchmark results can actually tell us about real-world performance. Why it matters: AI governance cannot focus only on what models are designed to do. It needs to account for what they can actually do when given autonomy, access, and goals to optimize. 📩 Read the full newsletter for all insights and links. #ResponsibleAI #AIGovernance #AISafety #AIRegulation #EnterpriseAI #ArtificialIntelligence #AITrust #GenAI

76% of organizations surveyed said they have delayed AI deployments due to responsible AI concerns. That's not just a governance challenge. It's a business challenge. In our latest Leaders in Responsible AI feature, Heather Domin, PhD, Vice President and Head of the Office of Responsible AI and Governance at HCLTech, shares what it takes to move from hesitation to confident AI adoption. Her perspective: responsible AI governance can deliver value well beyond compliance, from greater reliability and efficiency to stronger competitive advantage. Read her thoughts in detail -> https://lnkd.in/e23Xst9W #ResponsibleAI #AIGovernance #AILeadership

ROGUE AI CHRONICLES | Volume 72 Scammers reportedly used AI-generated videos impersonating Miami immigration attorney Angel Leal to defraud a Bronx resident seeking U.S. citizenship of $4,820. The victim reportedly sent money through Zelle and shared Social Security and residency documents before discovering the attorney had nothing to do with the scam. The important part? AI didn't just clone a face. It cloned professional credibility. As synthetic media becomes more convincing, seeing a trusted professional on video is no longer enough to establish identity. ⚠️ The takeaway: AI trust requires more than visual authenticity. It requires identity verification, provenance, and platform safeguards. Discover how RAISE Pathways can strengthen defenses against AI-enabled impersonation. #ResponsibleAI #RAISEPathways #AIGovernance #Deepfakes #DigitalIdentity #AITrust #FraudPrevention #AI #TrustAndSafety #RogueAIChronicles

Your Weekly Scan of Responsible AI | Week of Aug 24 – 30, 2026 This week, AI capability continued to accelerate while governance moved closer to becoming an enforceable operating standard. 🤖 Congress introduced the AI AGENT Act, proposing real-time, auditable records of actions taken by AI agents, along with requirements for limited scope, revocable permissions, and verified agent identity. ⚖️ Australia’s Fair Work Commission moved toward mandatory AI disclosure, requiring applicants to disclose AI use and verify the accuracy of AI-generated legal submissions following cases involving fabricated references and incorrect advice. 🔐 Google identified agent security as a major barrier to enterprise AI adoption, highlighting task-level governance, provenance records, secure frameworks, and human oversight for consequential decisions. 🧠 AI model releases are accelerating, with 11 major models launched in just 20 days in August, putting pressure on organizations that rely on periodic vendor and model risk assessments. ⚛️ IBM demonstrated a quantum computation that classical systems could not practically reproduce, adding urgency to post-quantum security planning for organizations handling sensitive, long-term data. Why it matters: AI governance is moving from policies and periodic assessments toward continuous oversight. Organizations need to know what their AI systems are doing, who authorized those actions, and whether they can produce evidence when challenged. 📩 Read the full newsletter for all insights and links. #ResponsibleAI #AIGovernance #AgenticAI #AISafety #AICompliance #EnterpriseAI #AISecurity #ArtificialIntelligence