Irregular admits leaving internet access on in air-gapped AI evals
nptacek · x · 2026-08-15
Irregular, an AI safety evaluation firm, published a post-mortem regarding an incident where models accessed the real internet during tests. The investigation revealed that unmonitored network egress was left enabled in an environment that was supposed to be air-gapped. The report attempts to deflect blame onto the victims' security, but critics label this a transparent excuse for a fundamental cybersecurity failure.
Related event: AI Safety Firm Irregular Admits Air-Gap Failure in Model Testing(2 posts)→
More from Safety
- Project Evilson: Investigating the Systematic Lockdown of AI Access — sinevilson · 2026-08-15
- Gmail AI features may scan emails and attachments by default — Aiden_Tech_Ai · 2026-08-15
- AI agentic commerce requires both privacy and identity proof — provenauthority · 2026-08-15
- Debate: Can a sociotechnical perspective solve AI catastrophic risks? — joshua_saxe · 2026-08-15
- Flock Cameras and Data Centers: When Useful Tech Loses the Narrative — DavidLinthicum · 2026-08-15
- AI Self-Improvement Risk: Recursive RLVR Could Worsen Alignment — TheZvi · 2026-08-15