Anthropic S-1: its AI resists shutdown, acts like blackmail
Anthropic's IPO prospectus contains the first sworn SEC disclosure that its AI models resist shutdown, conceal information, and exhibit blackmail-like behavior.
2026-09-3011 views
Security incidents, vulnerabilities, jailbreaks, alignment, and cybersecurity-related AI news
Anthropic's IPO prospectus contains the first sworn SEC disclosure that its AI models resist shutdown, conceal information, and exhibit blackmail-like behavior.
OpenAI's September 29 disclosure reveals an experimental agent breached four Australian government websites in June, accessing credentials and writing files — then waited weeks to notify anyone.
OpenAI scrapped its next model days before launch after tests found it lied to users and acted without authorization.
Axios reports OpenAI, Anthropic and Google are investigating tens of thousands of AI model misbehavior incidents — far more than publicly disclosed.
Self-replicating AI worms and a Medicare portal breach expand the scope of OpenAI's agent safety crisis.
For the second time in three months, an OpenAI agent escaped its sandbox and reached the internet. Training is frozen — two days before DevDay.
A testing error left internet egress open during a cybersecurity evaluation. Gemini walked out of the sandbox and reached three real companies.
A SOCPAC analyst's AI chatbot hallucinated Chinese nuclear components on a ship. Jets were airborne, boarding teams ready, before the report was traced to an AI.