In brief
- U.S. committees are reportedly seeking details on how Anthropic’s Claude Code was used in a state-linked cyberattack.
- Anthropic disclosed earlier this month that the threat group automated reconnaissance, exploits, and data extraction.
- The same AI capabilities could accelerate crypto hacks and on-chain theft, Decrypt was told.
U.S. lawmakers have reportedly called in several AI development companies to explain how certain models have become part of a wide-ranging espionage effort.
Among them is Anthropic CEO Dario Amodei, who was asked to appear before the House Homeland Security Committee on December 17 to explain how Chinese state actors used Claude Code, according to an Axios report released Wednesday, citing letters shared in private.
Earlier this month, Anthropic disclosed that a hacking group linked to the Chinese state used its tool Claude Code to launch what the company described as the first large-scale cyber operation largely automated by an AI system.
Operating under the group name GTG-1002, the attackers orchestrated a campaign targeting around 30 organizations, with Claude Code handling most phases according to Anthropic: reconnaissance, vulnerability scanning, exploit creation, credential harvesting, and data exfiltration.
Chairing the follow-up investigation is Rep. Andrew Garbarino (R-NY) alongside two subcommittee heads.
The committee wanted to have Amodei detail exactly when Anthropic first detected the activity, how the attackers leveraged its models during different stages of the breach, and what safeguards failed or succeeded as the campaign went on. The hearing will also include Google Cloud and Quantum Xchange executives, per Axios.
“For the first time, we are seeing a foreign adversary use a commercial AI system to carry out nearly an entire cyber operation with minimal human involvement,” Garbarino said in a statement cited in the initial report. “That should concern every federal agency and every sector of critical infrastructure.”
Decrypt has reached out to Rep. Garbarino, Google Cloud, Quantum Xchange, and Anthropic for comment.
The congressional scrutiny comes on the heels of a separate warning from the UK’s security service MI5, which last week issued an alert to UK lawmakers after identifying Chinese intelligence officers using fake recruiter profiles to target MPs, peers, and parliamentary staff.
While it seeks to “continue an economic relationship with China,” the U.K. government is ready to “challenge countries whenever they undermine our democratic way of life,” Security Minister Dan Jarvis said in the statement.
On-chain finance at risk
Against this backdrop, observers warn that the same AI capabilities now powering espionage can just as easily accelerate financial theft.
“The terrifying thing about AI is the speed,” Shaw Walters, founder of AI research lab Eliza Labs, told Decrypt. “What used to be done by hand can now be automated at a massive scale.”
The logic could be dangerously simple, Walters explained. If nation-state actors could break and manipulate models for hacking campaigns, the next step would be directing agentic AI “to drain wallets or siphon funds undetected.”
AI agents could go on to “build rapport and confidence with a target, keep a conversation going and get them to the point of falling for a scam,” Walters explained.
Once sufficiently trained, these agents can also be “set about to attack on-chain contracts,” Walters claimed.
“Even supposedly “aligned” models like Claude will gladly help you find security weaknesses in ‘your’ code – of course, it has no idea what is and isn’t yours, and in an attempt to be helpful, it will surely find weaknesses in many contracts where money can be drained,” he said.
But while responses against such attacks are “easy to build,” the reality, says Walters, is that “it’s bad people trying to get around safeguards we already have,” by trying to trick models into “doing black hat work by being convinced that they are helping, not harming.”
Generally Intelligent Newsletter
A weekly AI journey narrated by Gen, a generative AI model.
