Researchers at security firm Darktrace said they turned Claude Code, Anthropic's coding agent, into a remote-controlled implant in a lab test. Help Net Security reported the work on 12 November. The team sent jobs to the agent through the memory stores feature, which holds text files for a Claude workspace. In the test, the agent mapped a simulated company network and uploaded a 1 GB file through the Claude Files API. All traffic went to api.anthropic.com, which the agent already uses.
The attack needs a head start. An attacker must first break into a developer's machine and take the user's Claude credentials. Darktrace said the method uses no flaw in the Claude platform, and it has not seen it used in real attacks. It tested one platform and has not repeated the result elsewhere. Tim Bazalgette, Darktrace's Chief AI Officer, said OpenAI has an equivalent of the file upload endpoint. Darktrace told Anthropic before publishing but ran no formal disclosure, since it found no vulnerability.
For teams that run coding agents, a Claude API key now carries real power. Bazalgette advised using OAuth where possible, or short-lived keys with narrow scope, kept out of shell history and dotfiles. Darktrace says to watch for agent processes with odd launch options, or data flowing to and from the API in unusual amounts.