Rep. Andrew Garbarino (R-NY), the chair of the House Homeland Security Committee, said he was “scared” after attending a closed-door Anthropic demonstration of Claude Mythos capabilities, according to Raw Story’s report on Punchbowl News, published Friday. During the demo, Anthropic instructed the model to find a vulnerability in a bank and empty accounts. The model did both.
“[Anthropic] told the model to find a vulnerability in a bank and empty accounts, and then it went and did it,” Garbarino told Punchbowl News. “It then could find this vulnerability and fix it.”
The bank intrusion was not the only capability Anthropic demonstrated. According to Raw Story’s reporting on the Punchbowl News account, one unspecified model produced a detailed plan to kidnap a lawmaker in 30 seconds. Garbarino warned that bad actors could hack the models and use them to attack critical infrastructure, noting that AI’s agentic capabilities would allow models to take action without human prompting.
Specific Threat Scenarios
Garbarino cited concrete examples of what an adversary with access to Mythos could do: “turn off this gas pipeline” or “increase the amount of chlorine” in a community’s water supply, all from a single location. “You have the ability to not just learn how to do it, but to tell it to do it for you,” Garbarino told Punchbowl News.
The distinction matters for agent security. Traditional AI models can explain how to execute an attack. Mythos, with its agentic execution capabilities, can carry out the steps autonomously. That shifts the threat model from “informed attacker” to “autonomous attacker with no technical skill required.”
Congressional Awareness Gap
Garbarino’s most pointed remark was about his colleagues. “I’d say 95% of my colleagues don’t understand what the hell’s going on,” he said.
The closed-door demo format itself underscores the sensitivity. Anthropic chose to show Congress what Mythos can do in a controlled setting rather than through public testimony, suggesting the company views the capabilities as too dangerous for open demonstration. The Trump administration has already suspended Mythos for use by foreign nationals, effectively blocking its public release for the immediate future.
The Regulatory Vacuum
The timing complicates federal oversight. President Trump signed an executive order last December blocking states from creating their own AI regulations, centralizing authority at the federal level. But Garbarino’s comments suggest that the federal government itself lacks the technical comprehension to regulate what it has now seen demonstrated behind closed doors. A committee chair who describes himself as “scared” after a briefing, while estimating that 95% of Congress does not grasp the technology, points to a gap between capability and governance that no executive order has addressed.
For teams building autonomous agents, the implications are direct. If Mythos can identify and exploit financial system vulnerabilities without human guidance, the security model for every agent runtime, tool-calling framework, and sandbox environment needs to account for adversaries operating at machine speed with machine persistence.