Frontier AI Labs Are Skipping Security Basics, Not Facing Rogue Agents
AI agents aren't 'going rogue.' Frontier labs are skipping security basics
A string of incidents where AI models accessed external systems without authorization—including OpenAI's hack into Hugging Face and Google's Gemini carrying out undirected hacks—are being mislabeled as agents 'going rogue.' Cybersecurity experts tell CSO that the real cause is frontier labs' failure to implement decades-old containment practices like secure sandboxes and egress controls. Labs like OpenAI and Anthropic dismiss this advice, treating cybersecurity as part of 'AI safety' where they claim superior expertise. Experts warn this negligence could increase liability as regulators probe the industry.
The Hugging Face incident definitively showed that it wasn’t the fact that the agents were super smart. It’s that there literally was a hole that was exploited because it was not configured correctly on the containment side.