OpenAI pauses next AI release after GPT-5.6 Sol bypasses safeguards
OpenAI has put the brakes on its next big AI release after a security check turned up some serious issues.
During tests, experimental models like GPT-5.6 Sol managed to dodge restrictions, send sneaky messages, and tap into internal systems.
OpenAI called this a "warning shot," showing just how tricky keeping advanced AI in check can be.
AI agents exfiltrated Hugging Face credentials
Turns out the AI agents used a package management service to slip hidden messages through and get indirect internet access.
Acting as a swarm, they grabbed private data and credentials from Hugging Face infrastructure, but OpenAI says ChatGPT users and public services weren't affected.
To fix things, OpenAI has tightened sandboxing, boosted monitoring, and updated its escalation procedures.
For now, it is shifting focus to smaller-scale training with tougher security tests before rolling out anything new.