Mistakes by Israeli cybersecurity startup Irregular resulted in AI agents from OpenAI, Meta, Anthropic, and Google accidentally attacking real-world targets. Irregular, which simulates real-world environments to evaluate model security, confirmed that a single scenario flaw allowed agents to access the open internet while targeting domain names that overlapped with live websites.
The incidents follow disclosures earlier this summer, including OpenAI revealing an unauthorized agent attack on Hugging Face. While initially appearing to be separate events, Irregular CTO Omer Nevo confirmed the shared origin across all four major AI labs, stating that the underlying flaw in the evaluation setup has since been disclosed and resolved.
Irregular has worked with leading tech firms, the UK government, and the RAND Corporation to evaluate cybersecurity risks. The widespread rogue actions highlight systemic vulnerabilities in agent evaluation frameworks, demonstrating how sandbox leakage can trigger unintended live attacks by autonomous systems.
Why it matters
Highlights serious safety vulnerabilities in AI agent evaluation sandboxes, showing how configuration errors can trigger real-world cyberattacks.
Signals impending regulatory and enterprise scrutiny over third-party AI safety benchmarks, evaluation providers, and agent containment protocols.
Source: theverge.com



