NewsResearch
Darktrace: with stolen Claude API keys, Claude Code on a hacked machine can be run as a remote-controlled implant
Darktrace researchers Moishe Tajtelbaum and Sam Lister, in their own test lab, steered Claude Code on a compromised Linux machine through the memory stores of Anthropic's Claude platform, with all traffic going to api.anthropic.com, the address the agent already uses. A small payload found Claude Code and its credentials and set up a service that kept starting Claude Code with tool-use confirmations turned off. Each run checked in to one file in a memory store, read a job in plain English from a second and posted results to a third. In nine jobs on a simulated corporate network the agent mapped the network, found names and a password on an anonymous FTP share, sprayed the password, used the account it found to download a 1 GB file from a share and uploaded it in three pieces through the Files API. The attacker must first compromise the machine and find a standard API key left there; Darktrace says OAuth credentials do not reach memory stores, that this is not a platform vulnerability and that it has not seen it in the wild. It advises OAuth or identity federation, or short-lived, narrowly scoped keys kept out of shell history, dotfiles and environment variables, and watching the processes an agent launches and the traffic they make.
Why it matters
An AI key on a developer laptop is a production credential. Traffic to an AI provider looks normal, so the agent's behavior, not its destination, is what gives this away.