Lab test turns Claude Code memory into a C2 channel
Original title: Hackers Could Hijack Claude Code Memory to Control AI Agents and Steal Sensitive Corporate Data
With stolen API credentials, an operator queued jobs and pulled a 1 GB file through Anthropic’s own APIs.
Darktrace says attackers who already control a device can turn Claude Code into a command-and-control implant by abusing memory stores, which are text files scoped to a Claude Console workspace. In a lab Active Directory setup, a shell payload stole API and OAuth credentials and relaunched Claude Code harnesses that called Claude Sonnet 4.5. The agent ran discovery, password-sprayed an FTP password over SMB, and uploaded a split 1 GB file through the Claude Files API. Darktrace says the test used no Claude platform vulnerability and advises watching unusual harness behavior against baselines.
Unconfirmed: Darktrace’s results are from a controlled lab, not reported attacks on customers.
- Heartbeat, input and output files carry check-ins, jobs and results.
- Traffic uses the same inference endpoint, so it resembles normal agent activity.
- The agent found FTP, SMB and SSH on 192.168.1.2 and the account greg.simon.
- Nine jobs included ip, ss, nc and nmap, then an SMB share listing.
- Named stealers QUIETVAULT, s1ngularity, SANDWORM_MODE and DUSTMAKER lack a live feedback loop.
Impact: Co-opted agent runs commands and exfiltrates files with user privileges.
- 1.Correlate Claude Code harness process creation with network activity against behavioral baselines.
- 2.Watch unusual command-line arguments and odd data-transfer ratios to the AI platform API.
- ATT&CK
- T1528 Steal Application Access TokenT1071.001 Web ProtocolsT1046 Network Service DiscoveryT1110.003 Password SprayingT1567 Exfiltration Over Web Service
- Vendors
- Anthropic
- Products
- Claude Code
- Organizations
- Darktrace
- AI models
- Claude Sonnet 4.5
From the original article
2 min read at gbhackers.com
Attackers who already control a device can turn Claude Code agents into live command-and-control (C2) implants. They used the Claude platform’s own memory stores feature. Darktrace stresses that the work was done in a controlled lab and does not rely on a vulnerability in the Claude platform.
By Mayura Kathir · gbhackers.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.