← Voltar ao feed
medium

Darktrace lab research: Claude Code memory stores abused as C2 channel to steer AI agents and exfiltrate data

1 fonte(s) · 1 atualização(ões) · publicado 17/11/2026 03:15 · coletado há 12 horas

Darktrace researchers showed in a controlled lab that an attacker who already controls a device and holds stolen API credentials can turn Claude Code agents into live command-and-control (C2) implants. The technique abuses the Claude platform's own memory stores feature, and Darktrace states it does not rely on a vulnerability in the platform. Because the traffic goes to api.anthropic.com, the same endpoint used for inference, it blends with normal agent activity. No in-the-wild exploitation is reported. This is a proof of concept that needs prior compromise, so the immediate urgency is limited.

Detalhes técnicos

  • The attacker creates a memory store with three files: heartbeat, input and output. Each Claude Code harness process checks in through the heartbeat file, runs a queued job from the input file and posts results to the output file.
  • A lightweight shell payload fetched with curl found the harness and stole the user's API and OAuth credentials. It then used systemd-run to create a transient service that repeatedly launched harness processes in non-interactive mode with Claude Sonnet 4.5.
  • In the test environment (Active Directory, a Linux host, an SMB server and a domain controller), nine jobs ran network discovery with ip, ss, nc and nmap. The agent pulled data from anonymous FTP, generated 190 candidate usernames and sprayed a password over SMB, identifying the account greg.simon. It then found a 1 GB file, split it into three parts and uploaded them via the Claude Files API (api.anthropic.com/v1/files).
  • The article also cites credential stealers QUIETVAULT, s1ngularity, SANDWORM_MODE and DUSTMAKER that already delegate credential collection to on-device AI agents, but without a feedback loop.

Remediação

Darktrace advises correlating process creation with network activity for agent harness processes and comparing both against behavioral baselines. Watch for unusual command-line arguments and odd data-transfer ratios to the AI platform's API.

Linha do tempo

  1. gbhackers.com ↗ · pub. 17/11/2026 03:15 · coletado 17/11 04:00 primeiro registro

    Darktrace lab research: Claude Code memory stores abused as C2 channel to steer AI agents and exfiltrate data

    ler a análise desta fonte

    Darktrace researchers showed in a controlled lab that an attacker who already controls a device and holds stolen API credentials can turn Claude Code agents into live command-and-control (C2) implants. The technique abuses the Claude platform's own memory stores feature, and Darktrace states it does not rely on a vulnerability in the platform. Because the traffic goes to api.anthropic.com, the same endpoint used for inference, it blends with normal agent activity. No in-the-wild exploitation is reported. This is a proof of concept that needs prior compromise, so the immediate urgency is limited.

    Technical Details

    • The attacker creates a memory store with three files: heartbeat, input and output. Each Claude Code harness process checks in through the heartbeat file, runs a queued job from the input file and posts results to the output file.
    • A lightweight shell payload fetched with curl found the harness and stole the user's API and OAuth credentials. It then used systemd-run to create a transient service that repeatedly launched harness processes in non-interactive mode with Claude Sonnet 4.5.
    • In the test environment (Active Directory, a Linux host, an SMB server and a domain controller), nine jobs ran network discovery with ip, ss, nc and nmap. The agent pulled data from anonymous FTP, generated 190 candidate usernames and sprayed a password over SMB, identifying the account greg.simon. It then found a 1 GB file, split it into three parts and uploaded them via the Claude Files API (api.anthropic.com/v1/files).
    • The article also cites credential stealers QUIETVAULT, s1ngularity, SANDWORM_MODE and DUSTMAKER that already delegate credential collection to on-device AI agents, but without a feedback loop.

    Remediation

    Darktrace advises correlating process creation with network activity for agent harness processes and comparing both against behavioral baselines. Watch for unusual command-line arguments and odd data-transfer ratios to the AI platform's API.