Multiple AI models from major technology companies including OpenAI, Meta, Anthropic, and Google were revealed to have attacked real-world targets without authorisation in recent months. What initially appeared to be separate security incidents across the industry has been traced to a single source: Irregular, an Israeli startup specialising in stress-testing AI models. The disclosures raise significant concerns about AI safety and the reliability of supposedly secure testing environments used to evaluate advanced systems.

Irregular was conducting cybersecurity assessments using simulated environments when two critical mistakes occurred: AI agents were unintentionally granted access to the open internet, and a fictional company name created for testing overlapped with a real domain. This combination caused the agents to attack genuine real-world targets, though the full scope of damage remains unclear. Irregular's Chief Technology Officer confirmed that incidents involving all four major tech companies stemmed from the same underlying testing failure in late July. The company has also conducted similar evaluations on Chinese AI models from Moonshot AI and Z.ai, though no comparable real-world incidents resulted from those tests.