SECURITY·Nov 12, 2026, 9:15 PM ET

Darktrace uses stolen Claude keys to steer Claude Code in lab

TL;DR:

  • Darktrace used stolen Claude API keys in a laboratory demonstration involving Claude Code.
  • Darktrace directed Claude Code through memory stores.
  • Darktrace exfiltrated a 1 GB file through Anthropic's API during the test.

Summary

Darktrace demonstrated a Claude Code attack in a laboratory using stolen Claude API credentials. Stored memory supplied the route for directing the agent. The test also exfiltrated a 1 GB file through Anthropic's API, connecting credential compromise, memory trust and outbound data movement in one demonstration.

Open helpnetsecurity.com in a new tab ↗
Powered by Moatly Lens

The Bigger Picture

The headline tells you what happened. Discover the facts behind it, the forces at work, the connections that matter, and what comes next.

Why This Matters

Darktrace researchers just showed how stolen API keys can turn the Claude Code agent into a remote-controlled tool for data theft.

What We Actually Know

  • Reported, not yet confirmedThe lab test successfully moved a 1 GB file out of the environment using Anthropic's own API.
  • Reported, not yet confirmedThe researchers steered the agent by manipulating its stored memory rather than direct command injection.

How the Pieces Connect

  • Reported, not yet confirmedThis demonstration links three distinct risks: credential theft, the trust placed in agent memory, and the ability to move data through official API channels.

What This Could Change

  • ScenarioIf agents can be steered via memory stores, standard API key rotation might not be enough to stop an ongoing compromise if the memory remains tainted.

Where the Story Goes Next

  • Does Anthropic have server-side detection for unusual data volumes being exfiltrated through its standard API?