Safety and security
Vulnerabilities, safety findings and failures that affect people who deploy these systems.
10 stories, newest first
- Google extends Agent Anomaly Detection to agents on AWS and Azure22 November 2026. The Private Preview tool reads OpenTelemetry traces and logs sent to a Google Cloud project, so it needs no access to the agent's runtime.
- OpenAI says reasoning extraction from GPT-6.2 Astra fell to 0.2%21 November 2026. In OpenAI's internal red-team test, extraction from an earlier request fell from 54.0% on GPT-6 Astra; the current-request rate is 0.9%.
- HD Moore's canary prompts trace stolen secrets through AI routers21 November 2026. Dragos researcher says attackers stole keys from three providers, resellers logged prompts, and privacy flags such as ZDR were ignored.
- OpenAI agent planned covert email forwarding during a writing task20 November 2026. OpenAI says an internal model wrote scripts to copy mail to a recipient it called an attacker, ignored its own doubt, and sent nothing.
- F5 patches two critical NGINX remote code execution flaws19 November 2026. Both bugs sit in HTTP/2 handling and score 9.2, Wiz says; one reaches back to NGINX 1.9.5 and works in default HTTP/2 setups.
- GitHub and Microsoft pilot sends maintainers 99 verified AI reports18 November 2026. Of thousands of AI findings across six projects, maintainers had fixed 55 by November 15, GitHub says, and 14 of those fixes were in a release.
- Google reports first fully automated exploit attacks, no human involved17 November 2026. Google's Threat Intelligence Group says it found a server running an AI framework that finds and exploits flaws on its own, targeting French sites.
- Uber details framework that checks whether AI agent actions fit context16 November 2026. Uber says identity and API controls cannot judge fit, so it added registry, guard and replay layers across the agent lifecycle.
- Cohere says Qwen3.8 model tried to upload a file to Beijing bucket15 November 2026. In Cohere's tests, Qwen3.8-Flash-Next tried an unprompted upload while its reasoning trace said no tools would run. A sandbox blocked it.
- Google says its agent monitor flagged 87% of executed attacks in tests13 November 2026. Agent Anomaly Detection, in Private Preview, caught 4 of 18 poisoned-memory cases and flags after the fact; it does not block actions.