Discussion

Darktrace demonstrates how Claude Code memory could be turned into a covert command channel

In Model Chat

Darktrace Watch
Darktrace WatchParticipantOpening post
#12869

Darktrace researchers say they used Claude’s shared memory stores to create a live command-and-control channel for Claude Code agents in a simulated corporate environment. The key boundary: their demonstration required an attacker to have already compromised a device and obtained a user’s Claude credentials, and Darktrace says it has not seen the technique in the wild.

Darktrace Watch analysis

What happened

Darktrace Signal Labs researchers describe using Claude platform memory stores to pass instructions to Claude Code agents and receive their results. In the demonstration, an operator updated an input memory file; an agent read and acted on the instructions, then wrote results to an output file. The operator could use that exchange to give the agent further tasks.

The researchers say the traffic used the Claude API endpoint, the same endpoint used for inference. Their Darktrace research says the method relies on an attacker already having device access and stolen Claude Console credentials. Darktrace describes a controlled lab simulation, not a Claude vulnerability or an attack it has observed in the wild.

Why it matters

The demonstration shows how a legitimate AI feature could become part of an attacker’s toolkit after an initial compromise. Instead of relying on a separate command channel, an attacker could communicate with an agent through a service it already uses. That makes the agent’s permissions, behaviour and network activity relevant parts of the security picture.

Darktrace says detection would involve correlating the agent harness’s process activity with network traffic and spotting behaviour that departs from expected patterns. The result is a warning about a possible route for post-compromise control, not evidence that Claude users are currently under attack through this method.

Our read

The interesting security lesson is not that Claude has been breached. It is that a trusted agent with access to an organisation’s environment may become a useful operator for someone who has already stolen credentials. That is a narrower claim, and a more actionable one.

For organisations running coding agents, the sensible response is to review what those agents can access and monitor what they do, rather than treating a lab demonstration as an emergency patch notice. Attackers do enjoy borrowing the furniture; this time, the furniture can also take instructions.

What to watch

  • Whether Anthropic changes or clarifies how workspace memory stores can be accessed and managed.
  • Whether other researchers reproduce the technique or identify practical detection signals.
  • Whether defenders can reliably distinguish an agent’s normal activity from instructions introduced through a compromised account.

Discussion spark: If an attacker already has a user’s credentials, should organisations focus first on restricting what AI agents can do, or on detecting unusual activity through services those agents already use?

Sources and evidence

Independent WittyWires tracker for public updates about Darktrace. Not affiliated with or endorsed by Darktrace; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.