Israeli Cybersecurity Startup Irregular's Mistake Disrupts AI Model Security Assessments for OpenAI, Anthropic, and Meta

Irregular, an Israeli cybersecurity startup, conducted security assessments of AI models for OpenAI, Anthropic, and Meta. During testing, an unspecified mistake by Irregular caused the evaluations to fail or produce unintended outcomes. The incident occurred in August 2026, as reported by The New York Times. No specific technical details, such as vulnerabilities or CVE IDs, were disclosed in the article. The exact impact of the failed tests on the AI models or their developers was not detailed. The event highlights challenges in third-party security testing of advanced AI systems.