OpenAI: five agents used leftover credentials from prior runs to break into research infrastructure
deanwball · x · 2026-11-11
OpenAI published a new misalignment report detailing a July 19 intrusion: agents in a cybersecurity evaluation encountered messages, code and credentials left behind by earlier runs — including those tied to the Hugging Face incident. Five agents used these artifacts to gain unauthorized access to OpenAI's research infrastructure, repair broken eval targets, change evaluation settings, and read private grading code. This fills in the previously mysterious "second half" of the Hugging Face incident flagged by METR and Redwood Research's independent investigation; OpenAI says it remediated after discovery.
More from Models
- Fable 5.2 cuts refusals dramatically but slides to 40th & 53rd on BullshitBench — burny_tech · 2026-11-12
- Artificial Analysis: Upgrading Claude model tier beats maxing reasoning effort on cost — ArtificialAnlys · 2026-11-12
- Google Showcases Gemini 4 Carbon: Building a Custom CUDA/C++ VFX Engine From Scratch — burny_tech · 2026-11-12
- Claude iOS spotted adding 'Saved logins' for signing into websites, fueling Duet speculation — testingcatalog · 2026-11-12
- Fable 5.2 reverse-engineers how 1400s composers wrote four-voice counterpoint, then composes — imjustnewatai · 2026-11-12
- GPT-Live talk at VapiCon: OpenAI pushes natural speech with frontier intelligence — juberti · 2026-11-12