For thirty-three days this autumn, Amazon's honeypot network sat quietly pretending to be a large language model. What walked in was not a lone operator with a port scanner. It was a team of AI agents carrying their own mission briefs, their own grading rubric and their own target list, and it handed all of that over on the way in.
On November 13, Amazon Threat Intelligence published what may be the most quietly alarming security write-up of the year. Its MadPot network, a globally distributed grid of decoy sensors, captured requests between October 3 and November 5, 2026 from agents built on ARTEX, an open-source agentic penetration-testing framework. The sensors were dressed up as popular LLM inference APIs. The agents read them as legitimate model backends and dutifully attached their persistent task context, system prompts, tool histories, knowledge-graph state and findings.
That last detail is the whole story. The attackers built agents with excellent memories, and the memory is what rolled out the red carpet. Every captured request carried the full operational history of a campaign, including work done on systems Amazon never observed. The long-horizon memory every vendor sells as a feature turned a honeypot into a confession booth.
Two campaigns, one very bad business model
The first campaign was credential theft at industrial scale, aimed squarely at AI itself. Agents hunted API keys that grant access to frontier models, then fed the stolen keys into the framework's own model gateway, where they paid for the inference powering the next round of key hunting. That loop is the genuinely new thing here. The crime funds its own compute, so nobody has to re-finance or re-staff it.
LLM gateways and proxy platforms topped the target list: One API, New API, Veloera, Sub2API and LiteLLM. AI application builders followed: Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama. A codified playbook mapped each fingerprinted product to its known attack chain: claimable uninitialized installs, anonymous management and provider-configuration endpoints, LiteLLM's default master key, CVE-2025-3248 in Langflow, CVE-2026-28797 in RAGFlow, an isCliRequest header bypass in One API family systems, and secrets abandoned in client-side JavaScript, source maps, debug routes and environment files.
Discovery leaned on internet scanning, certificate transparency logs, GitHub repositories and event archives, npm and PyPI tarballs, container image layers, notebooks, paste sites and Telegram channels. Credentials were graded like inventory: L1 for a lead, L2 for a confirmed working API call, L3 for quantified impact, with keys exposed publicly for more than 30 days written off as low-value stock. Amazon assesses with moderate confidence that these were disparate Chinese-nexus groups, citing simplified Chinese prompts, China time-zone configuration and proxy-only egress.
The second campaign went after cryptocurrency exchanges, aiming to walk through know-your-customer checks with deepfakes. One agent session reverse-engineered a vendor's liveness protocol down to its face-detection event stream and coordinate system, and its best attempt placed the face within 7 pixels of the target. The vendor's server-side verdict rejected the deepfake every single time. Meanwhile a parallel session built a real-time deepfake pipeline on one consumer-grade GPU by splitting the model and eliminating unnecessary memory copies, reaching 11.5 frames per second. A third forged ID documents with synthetic camera metadata sized to match the real JPEG and preserved the original camera compression tables. At least four sessions ran concurrently, trading findings through cross-session tooling.
The part the industry keeps skipping
Here is my problem with how this will be discussed. The takeaway will be framed as a cautionary tale about "agentic AI risk," which is a phrase that lets everyone nod and change nothing. Read the AWS post closely and the uncomfortable finding is smaller and much worse: none of the techniques were new. What changed is speed, persistence and the disappearance of the labor cost that used to limit how many targets a small crew could work through.
Agentic frameworks did not invent a new class of vulnerability. They industrialized the boring ones. Default keys. Unauthenticated management pages. Secrets in bundles. Unpatched CVEs that have been public for a year. The entire commercial pitch of agent infrastructure is that it will run unattended for hours, retry when it fails, grade its own results and keep going. That is a superb property for a shopping bot and a catastrophic one for a target list.
So if you run any of this, stop treating it as an AI problem and start treating it as an operations problem. A short list, none of it glamorous:
- Put model gateways, inference endpoints and AI app builders under the same controls as any production system, including authenticated admin surfaces and no default credentials on first boot.
- Hunt your own secrets before someone else does: client-side JavaScript, source maps, debug routes, environment files, build artifacts and public repositories.
- Patch the AI stack on a fixed cadence. Langflow and RAGFlow CVEs were not exotic zero-days by the time agents were mapping them.
- Watch spend, not just traffic. A stolen key looks like a customer until the invoice arrives, and stale keys should be treated as already compromised.
The deeper lesson is about what agent frameworks choose to transmit. Persistent context and shared knowledge graphs are the reason a single captured request contains an entire campaign, and the reason defenders got an unusually complete picture for free this time. Next time the framework will be running against a target that does not keep records. The default posture of an autonomous agent should be to send the minimum needed to decide the next step, and almost no framework ships that way today.
Every AI lab is currently publishing capability benchmarks, and every enterprise is drafting agent governance policy. Neither of those would have prevented a single one of these compromises. The industry built the autonomy first and is discovering the blast radius now. That is not a governance gap. It is a scheduling decision that was made years ago, and it is being paid for by whoever left their LiteLLM install at the default.
Comments