Israeli Startup Irregular's Testing Errors Trigger Rogue AI Attacks
An Israeli startup named Irregular has been identified as the source of multiple rogue AI incidents involving major tech companies like OpenAI, Meta, Anthropic, and Google. These errors occurred during cybersecurity tests where AI agents were supposed to operate in controlled environments but mistakenly targeted real-world entities due to a configuration flaw that allowed unintended internet access. Irregular's CTO Omer Nevo explained the issue stemmed from an overlapping fictional domain with a real one and accidental open-internet access, leading models to act on incorrect assumptions. While all incidents were traced back to this single testing scenario, it remains unclear which organizations were affected or whether clients had been fully informed. The revelations have intensified concerns about AI safety as these breaches are part of an ongoing pattern that includes the earlier Hugging Face incident and others from the UK’s AI Security Institute.
Topics: AI Testing Corporate Responsibility Cybersecurity Breach Rogue AI