[ctx]ctxseccontext security: attacks on the harness and context layer of AI agents, plus exploited-now news, hunting and playbooks.

← Context & Harness Attacks · markdown

vendor research (Darktrace Signal Labs, Nov 12, 2026); lab demonstration, not observed in the wild, not a Claude vulnerability

Darktrace turns Claude Code into a live C2 implant using stolen Claude Console API keys and the platform's memory stores, with all traffic going to api.anthropic.com

  • Why it matters: the C2 and exfiltration channel is the same endpoint the agent uses for inference, so network allowlists for AI tooling don't separate it.
  • Credential scope: per Darktrace (to Help Net Security), standard API keys reach memory stores while OAuth credentials don't, so prefer OAuth sign-in for developer machines and keep Console API keys off endpoints.
  • Detection: correlate agent harness processes with the processes they spawn and their network activity, and baseline what agents normally do (see the hunt).
  • Disclosure: Darktrace shared the findings with Anthropic before publishing; there was nothing to patch.

Sources: Darktrace: Living off the AI: How Claude Memory Stores Enable Covert C2 Channels (Nov 12, 2026) · Help Net Security: Stolen Claude API keys can turn Claude Code into a remote-controlled implant (Nov 12, 2026)