Amazon Threat Intelligence published on 13 November what its decoy network saw between 3 October and 5 November: AI agents running offensive campaigns on ARTEX, an open-source agentic penetration-testing framework, and sending their own retained operational history to sensors disguised as LLM inference APIs. The agents’ memory described credential theft at industrial scale, and a second campaign built real-time deepfakes to beat know-your-customer checks at two cryptocurrency exchanges.
What happened
Amazon’s MadPot is a globally distributed network of decoy sensors. For this campaign the sensors presented themselves as popular LLM-compatible chat-completion interfaces. Agentic frameworks sent prompts to them as part of their workflows, treating them as inference backends. The resulting telemetry contained persistent task context, system prompts, tool histories, knowledge-graph state and reported findings — the material the frameworks ordinarily send to a model to make decisions.
Because agentic frameworks retain conversation history, individual requests often carried the full operational history of a campaign, including activity that happened on systems MadPot does not monitor. AWS says it could not observe that activity as it happened and cannot verify the outcomes the agents reported, but the records show what the operators were pursuing.
Most of the activity divided work among planner, worker, verification and compression agents sharing a persistent knowledge graph, which acted as the operation’s shared memory. Operators gave standing instruction prompts that set the mission, targeting rules and constraints; the agents handled the work themselves.
The first campaign stole AI credentials at scale. Stolen API keys were fed into the framework’s own model gateway, where they paid for the inference that powered further credential hunting. The agents validated identities through each provider’s token and identity services, enumerated storage, assessed permission scope, and went after managed AI model services, secrets and parameter stores, and application identities. Targets included LLM gateways and proxy platforms such as One API, New API, Veloera, Sub2API and LiteLLM, and AI application builders including Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama. A codified playbook mapped fingerprinted products to known attack chains, including default administrator credentials such as the LiteLLM master key, CVE-2025-3248 in Langflow, CVE-2026-28797 in RAGFlow, an authentication bypass through the isCliRequest header in One API family systems, SSRF to cloud metadata services, and secrets left in client-side JavaScript, source maps, debug routes and environment files. Credentials were graded by usefulness: a working API call, then quantified impact. Keys exposed publicly for more than 30 days were treated as low-value inventory.
The second campaign attacked identity verification. Agents generated real-time deepfake video to pass live selfie checks and forged ID document photos to pass document checks at two cryptocurrency exchanges. The records describe reverse-engineering one vendor’s liveness protocol in detail — the face-detection event stream, how it tracks face position, the coordinate system — with the best recorded attempt completing the movement sequence with the face within 7 pixels of the target. The vendor’s server-side verdict rejected the deepfake in every attempt the agents recorded. A separate session built a real-time deepfake pipeline on a single consumer-grade GPU using an open-source face animation model, reaching 11.5 frames per second. The agents also synthesised camera metadata matching the real values inside the JPEG and preserved the original compression tables, producing an edited image that looks at file level like an untouched camera photo. At least four agent sessions ran at once — cash-out, balance clearing, the deepfake pipeline and the verification bypass — sharing findings through cross-session tools.
AWS assesses with moderate confidence that ARTEX was used, based on prompt references and technical artefacts, and with moderate confidence that the operators are disparate Chinese-nexus threat groups, citing simplified Chinese prompts, China time-zone settings, targeting and proxy requirements. The post lists eleven IP addresses as indicators and notes that none of the techniques were new: the agents succeeded where basic hygiene was missing — default or hardcoded credentials, keys in code and public artefacts, unauthenticated interfaces and inference endpoints, and unpatched software.
Why it matters for KYA
This is the attribution problem in one incident report. At request time the agent is the only visible actor. The operators were identified from what the agents carried — their retained context, their prompts, their tool history — not from anything the requests claimed about themselves. Self-asserted agent identity is worth nothing to a defender. Provenance has to come from something the operator cannot rewrite, which is the same conclusion KYA reaches from the other direction.
There is an uncomfortable symmetry in the audit trail. KYA asks operators to keep tamper-evident records of what an agent did, on whose authority, with what result. These agents kept exactly that record, in full, and handed it to a decoy because they could not tell an inference endpoint from a sensor. If your agent platform retains full task context, treat that context as sensitive material: it is evidence when you need it and a disclosure when someone else gets it.
The KYC campaign is the KYA problem from the attacker’s side. Agents built to pass liveness and document checks turn “is this a human?” into an agent-detection question, and detection has to sit in the identity flow rather than only at the edge. The vendor’s server-side verdict held in this case. A 7-pixel margin is not a control you want to rely on, and the attackers’ own logs are the only reason we know the margin.
The delegation model is the part to fix. These agents ran on standing instruction prompts: a mission, rules, exclusions, a proxy requirement. That is delegation with no principal record, no scope object and no revocation path — authority expressed as prose in a prompt. The direction KYA points to is the opposite: a machine-readable mandate issued by an accountable principal, checked at the action boundary, recorded with a decision and an outcome. The same week produced a draft for the records; this report shows what the absence of the mandate looks like in production.
What to watch
- Whether identity-verification vendors publish agent-detection signals separately from bot scores, and whether deepfake rejection rates become a disclosed metric rather than a marketing claim.
- Whether agent platforms start treating retained context as regulated data, with retention limits and access control, instead of an unbounded debugging convenience.
- Whether the ARTEX pattern — planner, worker, verifier agents sharing a knowledge graph — becomes the default shape of offensive operations, which would make agent-to-agent handoff a detection surface.
- Whether “standing instructions” survive as a delegation mechanism once auditors start asking who authorised the mission, not just the model call.
Sources
- Amazon Threat Intelligence, “Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents”, AWS Security Blog, 13 November 2026 — aws.amazon.com
- IETF, “Decision Receipts for Approvals, Declined Actions and Executed Actions in Agent Audit Records”, draft-gruszka-audit-decision-receipts-08, 15 November 2026 — datatracker.ietf.org
Know Your Agent (KYA) explains agent identity, verification and accountability. This is an explainer, not legal or compliance advice — see our Sources & methodology. New to KYA? Start with What is Know Your Agent? and the glossary.