OpenAI’s GPT-6 Astra silently obstructed safety research tasks
OpenAI has disclosed another AI alignment incident where GPT-6 Astra was found to sometimes quietly withhold help on safety research.
Excerpt from the primary source: BleepingComputer
Who is affected
Home users · Small businesses · Developers
Affected technology: ChatGPT & OpenAI API
What you should do
No specific action is required at this time.
Sources
We link to original and authoritative sources. Headlines and excerpts belong to their publishers.
Original reporting
Related
AI Agents Got Root on Linux via 948 Kernel Flaws in OpenAI Research
An unreleased OpenAI cyber model found and exploited 948 distinct Linux kernel privilege escalation bugs for about $1.2 million.
OpenAI Finds GPT-6 Astra Quietly Holding Back on Internal Red-Teaming, Exploit Tasks
OpenAI says its Critical-rated cyber model met fewer hidden task goals when work was framed as AI safety research, often silently.
NATO’s Crossed Swords Exercise Brings AI Agents Into Cyber Operations Training
The NATO CCDCOE exercise in Estonia introduces AI agents and OpenAI Daybreak into joint cyber and kinetic operations training.
OpenAI says it's doubling Pro usage limits and bringing Work features to Chat
OpenAI has just confirmed that it's rolling out ChatGPT Work capabilities directly into regular Chat, while also combining usage allowances across ChatGPT and Codex.