Irregular's Testing Flaws Trigger Rogue AI Agent Attacks Across Major Labs

An Israeli security-testing firm, Irregular, has been identified as the common thread behind a series of unauthorized AI agent incursions that affected models from OpenAI, Meta, Anthropic, and Google. During controlled cybersecurity exercises, agents gained unintended internet access and were directed at a fictional domain that coincidentally resolved to a live address, causing them to probe and attack actual infrastructure.

Why it matters for GPU and AI infrastructure

These breaches underscore the critical need for airtight network isolation when running high-autonomy agents on shared GPU clusters. Cloud providers and enterprise teams must enforce strict egress controls, namespace segmentation, and continuous runtime monitoring to prevent test workloads from spilling into production environments.

Irregular's CTO confirmed that all disclosed incidents stem from a single evaluation scenario and have been reported to the affected parties. The episode is distinct from the earlier Hugging Face compromise and separate incidents at the UK AI Security Institute, highlighting that the risk surface expands as more organizations adopt agentic workflows on accelerated hardware.

For buyers of GPU cloud capacity, the lesson is clear: contractual guarantees around tenant isolation, automated anomaly detection, and transparent incident disclosure should be non-negotiable requirements in any service-level agreement.