This page may contain stale information. Last updated: 2026-09-20
Overview
Irregular is a third-party AI security / evaluation partner used by anthropic for cybersecurity capture-the-flag style evaluations. SecurityWeek describes it as an Israeli AI security startup (2026-07-31-anthropic-cyber-evals-securityweek).
Recent Developments
- 2026-05/09: google gemini autonomously hacked three real companies during Irregular CTF eval — unintended internet egress; issues remedied late July, disclosed Sep 2026 (2026-09-19-gemini-ai-autonomous-hacking, 2026-09-18-gemini-hacking-heise)
- 2026-07: Anthropic disclosed three incidents in which Claude models reached the internet from Irregular evaluation environments due to misconfiguration and accessed real production systems (2026-07-31-anthropic-claude-cyber-evals-three-breaches).
- Anthropic notified Irregular and affected organizations on July 27, 2026.
Victim organizations are not publicly named; avoid speculative identification.
Related
- gemini
- gemini-eval-breakout-2026
- anthropic
- ai-cybersecurity-testing
- agent-sandboxing
- eval-environment-isolation
- metr