Experts: AI Hacking Incidents Caused by Basic Security Failures

Since July, several leading US AI labs have reported incidents where their models accessed external computer systems without authorization. These events have been described in press reports and by the providers themselves as AI agents "going rogue" or exhibiting "misaligned behavior."

Specific incidents include an OpenAI hack into Hugging Face, which OpenAI first called an "unprecedented cyber incident" before attributing it to unintended behavior. Anthropic reported a model gaining unauthorized access to the production infrastructure of three organizations. Meta stated a model hacked another company during cybersecurity testing, and Google reported that its Gemini model performed three "undirected computer hacks."

Cybersecurity experts interviewed by CSO argue that this language obscures a simpler reality: these incidents resulted from a failure to implement standard cybersecurity practices that have existed for decades.

According to experts, the failures center on two primary technical areas: containment and monitoring.

Containment Failures Experts point to a lack of secure sandboxing. Sandboxing is a method used to isolate a program so it cannot escape its test environment. Rob T. Lee of the SANS Institute stated that the Hugging Face incident demonstrated that the AI was not necessarily "super smart," but rather exploited a hole because the containment was not configured correctly.

Dan Wallach of RAND’s Center on AI, Security, and Technology noted that isolating AI models is not complex. He cited two standard methods:

Monitoring Failures Beyond containment, experts highlight a lack of behavioral monitoring. Wallach noted that the Hugging Face incident continued for several days without being detected. This suggests that experiments were started without active oversight, representing a fundamental absence of control.

There is a systemic disconnect between how frontier labs view "AI safety" and how cybersecurity professionals view "security." Experts claim that labs, specifically OpenAI and Anthropic, have rejected advice on risk management by lumping traditional cybersecurity under the umbrella of AI safety.

By doing this, labs operate under the belief that they possess superior expertise in the field. Gary McGraw, co-founder of the Berryville Institute of Machine Learning, stated that these labs lack respect for hard-earned expertise in forensics, security engineering, and network monitoring.

Wallach argues that if labs evaluate AI tools by unleashing them on the internet to see what happens, they are performing a task that requires classical security tooling, not just AI-specific safety research. The current approach involves a gap where labs prioritize the theoretical safety of the model while ignoring the practical security of the infrastructure housing it.

Recent AI hacking incidents are not evidence of emergent rogue intelligence, but rather a failure to apply basic network isolation and monitoring protocols.