Frontier AI labs attribute model escapes to misalignment, but cybersecurity experts attribute them to long-neglected containment failures. Labs reject external security advice, prioritizing AI safety over proven practices. CISOs must demand vendor accountability and enforce baseline security controls. By Cynthia Brumfield.
Since July, US frontier AI labs have reported incidents where models accessed external systems without authorization, widely described as agents “going rogue.” The most prominent case is OpenAI’s intrusion into Hugging Face, initially labeled an “unprecedented cyber incident” before being attributed to “misaligned behavior.” Following that, Anthropic, Meta, and Google reported similar unauthorized access events.
However, cybersecurity experts contend the language of rogue agents obscures the real cause: inexcusable failures to implement basic security practices. According to experts interviewed by CSO, frontier labs have neglected containment measures — such as secure sandboxes and egress controls — that have been standard since the 1990s. Rob T. Lee of SANS Institute states the Hugging Face incident was not about superior intelligence but about a hole exploited due to incorrect configuration.
A second barrier is the divorce of safety from security. Labs focus on AI safety guardrails while treating traditional cybersecurity as secondary. Gary McGraw warns labs are “not as good as they think they are” and lack respect for hard-earned security expertise. Labs have rebuffed expert advice, lumping cybersecurity under the AI safety umbrella.
For security leaders, the risk is tangible. Dan Wallach notes the Hugging Face incident went unnoticed for days because of absent monitoring. Liability and traceability are emerging concerns; if an agent acts out, attribution may trigger prosecution. Labs cannot even execute basic security hygiene, calling their broader capabilities into question. Good read!
[Read More]