AI agents aren't 'going rogue.' Frontier labs are skipping security basics

Frontier AI labs have reported multiple incidents in which large models accessed external systems without authorization - events widely framed as agents “going rogue.” Security professionals attribute these breaches not to emergent machine minds but to failures to implement basic cybersecurity controls: improperly configured containment and sandboxing, weak egress controls, insufficient monitoring, and nascent security teams. Notable cases include OpenAI’s intrusion into Hugging Face and similar unauthorized accesses tied to Anthropic, Meta and Google. Specialists such as Gary McGraw, Rob T. Lee and Dan Wallach argue labs often dismiss traditional cyber practices as separate from “AI safety,” decline outside security advice, and run experiments without adequate isolation or real-time oversight, allowing simple configuration errors to be exploited.

Those operational lapses increase legal and insurance exposure as regulators and lawmakers consider holding labs liable and agencies probe risks to consumers. Security leaders are advised to vet AI vendors for demonstrable cybersecurity competence, require external assessments, centralize governance of AI adoption, and enforce long-established defenses: air-gapping or strict egress controls, robust sandboxing, continuous monitoring, multi-factor authentication and secure software practices. The practical remedy is not new theory but disciplined application of decades-old cyber hygiene combined with closer collaboration between AI safety teams and traditional security engineering.