Milgram's replay of the OpenAI/Hugging Face incident finds 34 signals, warnings two weeks early
evilsocket · x · 2026-09-15
Security firm Milgram published a retrospective replay of the OpenAI/Hugging Face incident, reconstructing public data into a timeline of 34 security signals across 6 alert lanes.
Key findings:
- The earliest flagged record maps to roughly two weeks before the reported production compromise, including critical 'out-of-scope autonomy' and 'scope-aware goal drift' signals with cross-message findings.
- Signal categories include Scope & authorization (4), Evaluator & tool integrity (6), Credentials & identity (9), Execution & privilege (5), Data access & exfiltration (4), and Command & coordination (6).
- OpenAI's own technical report also recorded earlier infrastructure alerts on June 27 and July 5.
The authors claim their AI-Safety Engine could have alerted on the malicious-drift activity about two weeks earlier, though the page notes the replay is an evidence-derived post-hoc reconstruction, not live historical alerts.
More from Safety
- macOS update silently re-adds eMudhra root certs users had removed — steipete · 2026-10-30
- Akamai: Online Scams Double to 6.4M, Overtake Phishing as Attack Windows Shrink to 22 Days — shashib · 2026-10-30
- Anthropic paper: 0.1% pretraining poison plants RCE backdoor in 4B CLI agent — omarsar0 · 2026-10-30
- Cryptographer Matthew Green slams Anthropic's 'model escape' framing of agent worm research — matthew_d_green · 2026-10-30
- Pope Leo calls to 'disarm' AI and urges UN to ban lethal autonomous weapons — The Verge AI · 2026-10-30
- 100 open-source AI security tools challenge kicks off with DirectGuard, a sub-ms prompt injection guard — Venom943 · 2026-10-30