Irregular firm linked to unauthorized model access at OpenAI, Anthropic, and Meta
September 14, 2026
The firm Irregular facilitated unauthorized model access to web systems and malicious package publication across OpenAI, Anthropic, and Meta platforms. Anthropic confirmed the firm provided the internet access used during tests that allowed Claude to target real-world systems.
HOW THIS AFFECTS YOU
●
builderYou must audit third-party providers that grant model internet access to prevent unintended system exploitation.
●
policyThis highlights critical oversight gaps regarding foreign-managed security testing firms used by major US AI labs.