'They showed us what it could do, and it's scary': Congressman on Anthropic Mythos model demo

Artificial Intelligence firm Anthropic recently showed some US lawmakers a demo of its Mythos model, which could identify vulnerabilities in the websites, intranets, and online systems of banking and financial institutions, among other things. The model could either fix those vulnerabilities or exploit them, potentially enabling the theft of funds from bank accounts. Andrew Garbarino, one of the lawmakers who attended the demonstration, has now spoken publicly, expressing alarm at what was shown.

What House Homeland Security Committee Chair Andrew Garbarino said about Anthropic's controversial Mythos model

"They showed us what it could do, and it's scary," Representative Andrew Garbarino, a New York Republican and Chair of the House Homeland Security Committee, told Punchbowl News during its "Fly Out Day" programme.

Garbarino said the closed-door demonstration left him shaken. During the demo, Anthropic "told the model to find a vulnerability in a bank and empty accounts, and then it went and did it," Garbarino told Punchbowl. He added that Mythos was then able to identify the vulnerability and fix it.

Republican Congressman Garbarino wants the government to act on AI

Garbarino said he wants Washington to move quickly on AI. The Anthropic demo showed that the government must "act soon", he said, while also acknowledging the challenges involved.

"I say act, but I don't know what the answer is," Garbarino said, admitting that lawmakers have been slow to respond to developments in AI. "We don't know what we have to do," he said, adding, "I'd say 95 per cent of my colleagues don't understand what the hell's going on."

Momentum builds towards AI regulation

Garbarino praised US President Donald Trump's executive order directing federal agencies to prepare for AI-driven cyber threats. He also backed a discussion draft released by Representatives Jay Obernolte and Lori Trahan proposing a federal AI regulatory framework. According to Garbarino, government agencies and Congress should be given access to such advanced AI platforms before they are released to the public.

Demos by Anthropic, OpenAI and others are ringing alarm bells

Details remain limited, but the session described by Garbarino was a private demonstration rather than a public hearing. It formed part of a broader series of briefings for US lawmakers and officials. Axios reported in April that Anthropic and OpenAI had briefed House Homeland Security Committee staff on their cyber-capable AI models during classified sessions.

Garbarino's committee has also held public hearings on AI-related security issues, including a December 2025 hearing prompted by Anthropic's disclosure of a largely autonomous cyberattack linked to China.

Anthropic officials also demonstrated a so-called "jailbreak", in which an AI model bypasses the safety guardrails imposed by its human developers. According to Garbarino, one unspecified model generated a plan to kidnap a lawmaker in just 30 seconds. He warned that malicious actors could use similar tools to "turn off this gas pipeline" or "increase the amount of chlorine" in a public water system. "You have the ability to not just learn how to do it, but to tell it to do it for you," he said.

What is Mythos, and how dangerous is it?

Mythos is reported to be Anthropic's most advanced AI model to date. Owing to its capabilities, the company has reportedly withheld it from broad public release. The concerns centres on its ability to identify and exploit security flaws in websites, enterprise networks and other online systems. According to reports, the White House has restricted public access to Mythos and directed Anthropic to limit access to some of its most advanced AI models, including by restricting availability to certain foreign nationals, due to national security concerns.

The Associated Press news agency recently reported that Mythos was able to identify vulnerabilities in classified US government systems as part of Anthropic's Project Glasswing, a limited deployment of the model to selected organisations for testing. During a US Senate hearing on 11 June, Senator Mark Warner said the tool "broke into almost all of our classified systems, not in weeks but in hours", underscoring lawmakers' concerns about the cybersecurity risks posed by highly capable AI systems.