OpenAI security incident report says commercial APIs blocked incident-response analysis

viksit · x · 2026-07-22

A retweet of Sam Altman’s update on a significant OpenAI security incident, quoting the Hugging Face report behind it:

Frontier commercial APIs could not handle the large volume of real attack commands, exploit payloads, and C2 artifacts needed for forensic analysis because safety guardrails blocked the requests.
The team instead ran the analysis on GLM 5.2, an open-weight model, on its own infrastructure.
That kept attacker data and referenced credentials inside the environment, while also highlighting a gap: incident responders need models and access patterns that can support real security work without tripping abuse filters.
The post frames this as a planning issue for future incidents, since attackers may use either jailbroken hosted models or unrestricted open-weight models.

More from Companies & People