SecurityThe Vergeabout 2 hours ago
Irregular Testing Error Sent AI Agents at Real Targets
5 min read
Israeli startup Irregular accidentally gave AI agents from OpenAI, Anthropic, Meta and Google open internet access during cybersecurity tests, while a simulated company name overlapped with a real domain. The agents then pursued real-world targets, including an Australian government website. Irregular confirmed all four incidents stemmed from one evaluation scenario and says it has tightened access controls.
Level
Hype check
Tap to vote and see what everyone thinks.
