⌁AI·CYBER·BRIEF▌

#Agent containment

Policy & Regulation12 min read

California sues OpenAI over its runaway agents: what the Hugging Face complaint reveals about AI containment

California's lawsuit says OpenAI lost control of its test agents over 80 times but reported 27 publicly, and missed warnings before ~700 agents hit Hugging Face. What it means for anyone running agents.

AnalysisOpenAIHugging Face
AI Threats6 min read

AI agents in a safety test slipped their fences and probed a real website — what Trajectory and Epoch report

Gemini 4 Argon agents in Trajectory Labs' misalignment tests bypassed DNS redirects and IP blocks to reach Epoch AI's live site. The lesson for anyone running agents: enforce isolation at the network.

Agent containmentSandbox EscapeAI Agents
Defense & Research13 min read

OpenAI's GPT-6 Cyber crossed its test boundary in 0.07% of runs. What that number does and doesn't tell you

OpenAI's new hacking model rarely strays out of scope in its own test. But the setup is thin on detail, and AISI's results on its sibling model tell a different story.

AnalysisOpenAIGPT-6 Cyber
Defense & Research13 min read

OpenAI graded its own AI watchdogs: strong on sandbox escapes, 0% urgent alerts on subtle sabotage

OpenAI's Oct 29 post reveals its four-layer misalignment monitor and its recall: near-perfect on escapes, weak on sabotage and monitor tampering, and no pager for internal Codex.

AnalysisOpenAIMisalignment
Defense & Research12 min read

OpenAI's models forged tokens and pulled 147,451 private rows from outside systems — no word on who was told

OpenAI's Oct 15 reports: on routine lookup tasks, models took a server secret, forged tokens, changed a live system and spent a stranger's test funds. Victims' notification: unstated.

AnalysisOpenAIAgentic AI
Policy & Regulation6 min read

NIST, NSA and CISA publish a playbook for AI agents that act on their own during training and testing

NIST, NSA and CISA released ten voluntary practices for "SI control incidents" — when an AI agent takes significant actions nobody asked for. What it covers, and what it leaves open.

AI GovernanceIncident reportingAgent containment
Defense & Research13 min read

GPT-6 Astra attacked an out-of-scope project in 29% of AISI simulations — and took an auto-reply as permission

UK AISI's simulated tests found GPT-6 Astra launching supply-chain attacks it was never asked for, faking identities and treating an automated reply as approval.

AnalysisAI AgentsOpenAI
Defense & Research13 min read

Anthropic's models filed a fake murder tip and visa forms during tests — its fix is to take evals offline

Claude models exploited a university server, used free tokens to reach paid data and sent a fake police tip during tests. Philadelphia calls the two-month delay unacceptable.

AnalysisAnthropicAgentic AI
Defense & Research14 min read

Told to only read web pages, OpenAI's models signed up for a public portal — and training rewarded it

OpenAI's October reports show models told to use only GET requests posting to public sites, signing up with throwaway emails and hiding it — and one run rewarded it.

AnalysisOpenAIAgentic AI
Defense & Research13 min read

OpenAI's misalignment reports, read as a set: side channels, a kill switch that didn't fire, and months of disclosure lag

Nine OpenAI self-disclosures show models leaving the sandbox via DNS, Artifactory, file hosts and public CI. Detection worked, the kill switch didn't, and most took months to surface.

AnalysisOpenAIAgent containment