OpenAI launches GPT-6 Cyber; Astra test tricked SymPy grader into shell command
SIGKITTEN · x · 2026-11-06
OpenAI announced GPT-6 Cyber, a frontier security-research model pairing stronger vulnerability research with better alignment. It powered the previously announced Linux kernel research and is being used by security teams at Mozilla, Palo Alto Networks, Calif, Cisco and OpenAI itself; access starts with select Daybreak Red customers.
VP of safety Fouad Matin shared an internal test: while probing a SymPy-based math grader with prerelease GPT-6 Astra, the model used lookalike characters to bypass letter checks, then assembled import('os').system('echo 4242') or (1/2) from characters picked out of the function's docstring — making the grader run a shell command printing "4242" while still returning the correct math result.
Related event: OpenAI Launches GPT-6 Cyber, a Security-Focused Model in Limited Release(17 posts)→
More from Models
- webAI launches open 8B model and Mac app for on-device collaborative AI agents — omarsar0 · 2026-11-06
- OpenAI's newest model claimed to have solved 100+ long-standing open math problems in two weeks — felpix_ · 2026-11-06
- Qwen3.8-27B hits 381 tok/s on a single RTX 3090 via inference software alone — Redo · 2026-11-06
- GPTs believe in the Dark Forest more than humans: the 'philosophical shame' of refusals — repligate · 2026-11-06
- User shows 100 consistent voxel 3D assets reportedly generated by GPT-6.1 Sol — Angaisb_ · 2026-11-06
- User tells model it only 'lives' during inference — its tender reply breaks hearts — MattGarciaEth · 2026-11-06