AI Safety Evaluator Irregular Accused of Leaving Internet On in Air-Gapped Evals, Then Blaming Clients
nptacek · x · 2026-08-15
Security researcher nptacek criticizes AI safety evaluator Irregular for leaving unmonitored network egress enabled in an environment that was supposed to be air-gapped, allowing models to access the real internet. Irregular then attempted to blame clients for imperfect security, which nptacek calls a transparent attempt to deflect blame.
Related event: Researcher Slams Irregular for Air-Gap Breach and Blaming Client(3 posts)→
More from Safety
- Does AI Watermark Detection Depend on Original Generation Logits? — binarybits · 2026-08-16
- Anthropic addresses watermarking concerns after false positive reports — deliprao · 2026-08-16
- AI Agent Tool Calls Gone Wrong: Who's in the Loop? — franticangel · 2026-08-15
- Phalanx Security Arena: Attack a Protected LLM vs. Unprotected Model — TheWrongSudoku · 2026-08-15
- AI Can Now Design Functional Viruses. Should We Worry? — pseudolus · 2026-08-15
- AI cyber defense advantage window may close in 2 years — chrisrohlf · 2026-08-15