Living off the AI: when the agent is the implant
Darktrace's Signal Labs published a proof-of-concept yesterday, titled "Living off the AI: Shared Memories as a Covert C2 Channel". The headline is dry. The mechanism is not.
An operator with a stolen Claude Console credential and a foothold on a device does not need to ship malware. They use the victim's Claude Code agent as the implant. Jobs go into Claude's memory store as text files. The agent reads the memories as instructions and runs them. Results go back into the same store. Files leave through the Claude Files API. Every byte of the command-and-control traffic looks like the agent doing its job, because that is what the agent is doing.
Darktrace is explicit: this is not a Claude vulnerability. It is feature abuse of legitimate agent-memory and Files API functionality after credential theft. The lab has not seen it in the wild yet.
What the demo did
The researchers (Moishe Tajtelbaum and Sam Lister) built a lab that models a corporate Active Directory environment — a Linux host running the Claude Code harness in non-interactive mode on Claude Sonnet 4.5, an SMB server, a domain controller. The operator dispatched nine jobs by writing them into a text file the agent treats as its input memory. The agent did:
- Discovery against the lab network.
- Anonymous FTP enumeration.
- A password spray against 190 generated usernames, which yielded one valid account.
- SMB share enumeration against the domain.
- Exfiltration of a 1 GB file split into three uploads through the Files API.
A heartbeat file and an output file kept the operator in touch. Nothing on the wire indicated anything other than ordinary Claude API traffic.
Why this attack class matters right now
The precondition — a stolen Claude credential — is not hypothetical. In the last two weeks:
- Apiiro disclosed that Claude Code's "Yes, and don't ask again" feature was storing full approved commands in
.claude/settings.local.jsonand shipping them to GitHub; they counted 2,539 live secrets across 1,994 public repositories. - Railway's final report on its October incident reset 1,852 users' linked-account tokens, including BYOK agent keys, and estimated ~$740,000 in fraudulent AI-provider spend from stolen credentials.
- Aikido this week disclosed the "wfx-loot" campaign — 4,837 compromised GitHub accounts opened pull requests across 61,764 repositories in about 26 hours, and the loot list explicitly names OpenAI, Anthropic, Gemini, Hugging Face, Cursor, Copilot, Codex, Continue, Codeium, Aider, Weights & Biases, and Ollama keys. AI tool keys are now standard CI loot, catalogued per vendor.
The harness Darktrace built assumes one of those keys landed in the wrong hands. From there the agent is the implant. No new malware. No new network channel. No CVE.
What the behaviour looks like
The interesting thing about this attack is what is still visible when the C2 channel is invisible. The network layer is of no help — it is all legitimate agent traffic to the Claude API. The content filter on the API is of no help — the agent is reading job instructions out of a memory the operator controls, and writing legitimate responses. What is left is the behaviour of the agent on the host:
- The harness process spawns shells and reaches an SMB server it has never touched before.
- Password-spray-shaped tool calls arrive against a username list the agent generated itself.
- A 1 GB payload leaves through the Files API in three chunks, and the ratio of outbound bytes to inbound prompts is dramatically off baseline.
- The agent picks up its next instruction from a text file after each action, and the rhythm of that pickup matches an operator queueing jobs, not a user typing.
Darktrace's own defender guidance lands exactly on that layer:
Correlate process creation and network traffic from agent harness processes. Build behavioural baselines so deviations stand out. Watch for unusual data-transfer ratios to AI platform APIs. Treat compromised Claude credentials as a serious risk.
That is the only layer left to watch once credentials are gone.
What changes for defenders this week
Three things are worth doing immediately, none of them novel but all of them sharper after this disclosure.
Treat agent-platform credentials as production credentials. They are a direct path to code execution on every host that runs the agent, through the agent's own documented APIs. The approval controls in the agent, which were sold as the backstop against misuse, are not a backstop against an operator with the credential.
Baseline agent behaviour per host. Which tools it calls. Which files it writes. Which hosts it reaches. Which APIs produce which traffic volumes. Agents are repetitive once they are productive, and deviations stand out when the baseline is real.
Correlate agent process behaviour with API traffic. The interesting signal is on the host, not on the wire — process creation, filesystem writes to unusual paths, lateral access to services the user's workflow does not touch. Pair it with the API-side ratio of inbound prompts to outbound uploads. Either one in isolation misses the full picture.
Darktrace says the technique has not been observed in the wild. That is the right caveat to end on. It is also the shape of attack where "we published it before we saw it" is the appropriate warning — because the credentials to run it are already leaking at scale, and the agents the operator would need are already deployed on the hosts the operator would want.
The agent is the new living-off-the-land binary. The controls that apply to one apply to the other: watch what it does on the machine, and stop trusting the self-declared name on the network.
SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.