Last updated: 17 November 2026. Based on Darktrace's 12 November 2026 research and follow-up coverage. SeeHow this was checkedat the end.
Researchers at Darktrace showed that an attacker who already controls a developer's Linux machine and has stolen their Claude Console credentials can turn Claude Code into a remotely controlled implant. The command channel and the stolen data both travel through api.anthropic.com, a destination most companies allow by default. Nothing in this is a bug in Claude, and Darktrace says it has not seen it in real attacks. It still matters, because it moves the problem from "block the bad domain" to "notice the odd behavior of an approved tool."
Quick facts
| Item | Detail |
|---|---|
| Who | Darktrace Signal Labs (Moishe Tajtelbaum and Sam Lister), blog post dated 12 November 2026 |
| What | A lab proof of concept: Claude memory stores as a command channel, the Files API as an exfiltration path |
| Is it a vulnerability? | No. Darktrace says it "does not rely on a vulnerability in the Claude platform"; no CVE |
| Prerequisites | A compromised device and stolen Claude Console credentials |
| In the wild? | Not observed by Darktrace |
| Model used in the lab | Claude Sonnet 4.5, non-interactive mode |
| Anthropic response | None found. Help Net Security reports Darktrace shared the findings with Anthropic "as a courtesy" |
How the lab attack worked
-
Foothold and credentials. After compromising a Linux device, a small shell payload (run through
dash) looked for a Claude Code installation and sent the user's Claude API and OAuth credentials to a staging server. - The command channel. Using those credentials, the attacker created a memory store in the user's Claude Console workspace. Memory stores are collections of text files that agents read and update across sessions. Darktrace used three files: a heartbeat (check-ins), an input file (queued jobs) and an output file (results).
-
Persistence. The payload used
systemd-runto create a transient service that repeatedly launched Claude Code processes. Help Net Security reports they ran with--dangerously-skip-permissions, which turns off tool-use confirmation prompts. The first script was then discarded. - Tasking in plain English. Each process checked in, read the next job from the input file, ran it and wrote the result back. The operator gave goals, not commands.
- Results. Across nine jobs on a simulated Active Directory network, the agent found a host with FTP, SMB and SSH open, took files from an anonymous FTP share, generated 190 usernames, sprayed one password at the domain controller, found one valid account, located a 1 GB file on an SMB share, split it into three parts and uploaded them to the Files API.
Why it is hard to see
Darktrace's point is that the agent reads and writes the memory files through the same API endpoint it uses for inference, so the command traffic looks like normal Claude use. The process is an approved coding tool, the destination is a sanctioned provider, and no separate command server exists after the first stage. A firewall or proxy that allows api.anthropic.com sees nothing unusual. Help Net Security quotes Bazalgette, who is cited alongside Darktrace's findings, as saying "approved is not the same as safe, and only behaviour tells you the difference."
What this is not
- It is not a remote exploit. The attacker already needs a foothold and the credentials.
- It is not a prompt-injection or sandbox-escape result.
- It is one lab test on one platform. Darktrace says it has not reproduced the technique elsewhere. Help Net Security notes OpenAI has a comparable file-upload endpoint.
- The names and credentials in the published walkthrough are fictitious, so the IP address and account name in the write-ups are not indicators you can hunt for.
What to do
1. Treat AI platform keys like production credentials
The attack starts with stolen credentials. Per Darktrace and the Cloud Security Alliance's analysis of the research:
- Prefer OAuth or identity federation over static API keys. In the Help Net Security summary of Darktrace's findings, a standard API key could reach memory stores, while OAuth credentials could not.
- Where static keys are unavoidable, make them short-lived and limited to a single workspace, and keep them out of shell history, dotfiles and environment variables.
- Look for credentials on developer machines. The CSA note points at
~/.claude/.credentials.json,ANTHROPIC_API_KEYvalues and keys in shell history. - Rotate keys on any machine you suspect was compromised.
2. Block unattended permission bypass
The Claude Code documentation says bypassPermissions mode skips permission prompts and should only be used in isolated environments such as containers or VMs. It also documents a setting that prevents it, best placed in managed settings so users cannot override it. On Linux, the documented location is /etc/claude-code/managed-settings.json:
{
"permissions": {
"disableBypassPermissionsMode": "disable"
}
}
This reduces the chance of a silent, prompt-free agent on your developers' machines. It does not stop an attacker who can run their own binary or use the API directly, so it is one layer, not a fix. I did not test this setting; it is taken from the documentation.
3. Hunt for the process pattern
Darktrace's detection advice is behavioral: correlate agent process launches with network activity, compare against a baseline, and watch for unusual command-line arguments and abnormal upload-to-download ratios to the API. A few starting points on Linux (my suggestions, not from the report, and I could not run them in my environment, so adapt them first):
# Claude processes: who started them, how long they have run, and with what arguments
ps -eo pid,ppid,user,etime,args | grep '[c]laude'
# Transient services created with systemd-run
systemctl list-units --type=service --all 'run-*'
Things worth alerting on: a Claude Code process started by systemd-run or cron outside your approved automation, a Claude Code process running with --dangerously-skip-permissions on a developer laptop, and an agent process spawning reconnaissance tools such as nmap, nc or smbclient.
4. Review the workspace
- Check every Claude Console workspace for memory stores you do not recognize.
- Review Files API uploads for unexpected or large files.
- Remember that blocking
api.anthropic.comis unrealistic for teams that use it daily, which is why the monitoring above matters.
The bigger lesson
The same pattern applies to any approved service that can store data and run logic: a trusted destination hides a malicious conversation. The practical defenses are old ones, applied to a new kind of credential: least privilege, short-lived secrets, separate workspaces, and behavior-based detection. If your developers' laptops hold keys that can create storage and run an AI agent, those keys deserve the same protection as cloud admin tokens.
Related Dev Talk articles
- Cursor and Claude Code Extension Backdoor: What to Check
- Git Security Release CVE-2026-111760: What to Update
Frequently asked questions
Is this a Claude Code vulnerability?
No. Darktrace says it uses legitimate features and requires prior device compromise and stolen credentials. There is no CVE and no patch to install.
Has it been used in real attacks?
Darktrace says it has not observed the technique in the wild. The Cloud Security Alliance note expects key-stealing malware to add agent-tasking modules, but labels that an inference with no public evidence.
Do I need to stop using Claude Code?
Nothing in the research says so. Protect the credentials, restrict permission bypass, and monitor agent behavior.
Does it affect other AI coding tools?
Darktrace tested one platform and has not reproduced it elsewhere. Any provider API that combines text storage with inference could allow a similar pattern, depending on how credentials are scoped.
update log
- 17 November 2026: first published.
- To recheck on each update: any Anthropic statement or product change (alerts on new memory stores, controls for inference-only keys), reports of use in real attacks, and Darktrace's detection content.
Sources
- Darktrace: Living off the AI: Shared Memories as a Covert C2 Channel
- Help Net Security: Stolen Claude API keys can turn Claude Code into a remote-controlled implant
- Cloud Security Alliance: Living Off the Model API: Stolen Claude Keys as Agent Implants
- Cyber Security News: Claude Code's Trusted API Could Help Attackers Hide Malicious Commands and Data Theft
- Claude Code documentation: Configure permissions
- Claude Code documentation: Deploy managed settings
Originally published on DEV Talk.
Top comments (0)