The FBI says open-source AI models are increasingly matching frontier systems in finding and exploiting software flaws, signalling a new cybersecurity challenge that extends beyond proprietary AI.
The FBI has publicly acknowledged that open-source AI models are emerging as a significant cybersecurity concern, warning that offensive AI capabilities are no longer confined to frontier proprietary systems.
Speaking at the 930gov conference, Todd Hemmen, Deputy Assistant Director of the FBI’s Cyber Capabilities Branch, said Anthropic’s Mythos AI presents “future challenges for law enforcement.” He added that less capable models, including open-source AI, can identify and exploit vulnerabilities in ways similar to Mythos, broadening the threat landscape.
The warning comes as researchers demonstrated that an open-source 1.2-billion-parameter AI model running on a consumer laptop discovered real software vulnerabilities in production codebases in under four minutes at virtually no cost. Another study showed a prototype AI worm built entirely on free, open-source models successfully exploited 73.8% of a simulated enterprise network.
Anthropic’s Mythos Preview further highlighted the pace of AI-assisted vulnerability discovery by uncovering a 27-year-old OpenBSD TCP SACK flaw, a 16-year-old FFmpeg vulnerability missed by fuzzers after more than five million executions, and autonomously chaining Linux kernel vulnerabilities to gain root privileges. Across nearly 1,000 open-source repositories, the model identified thousands of potential high- and critical-severity vulnerabilities, with human experts agreeing with its severity assessments in 89% of reviewed cases.
“Mythos found vulnerabilities in some of the open-source code that is so ubiquitous — it’s in the vast majority of our most foundational code for things like operating systems, security, web infrastructure, encryption. It presents future challenges for law enforcement,” said Hemmen.
Anthropic has urged organisations to adopt AI-assisted defensive vulnerability research and accelerate patch cycles as autonomous exploit development continues to advance.
