Researchers usually reconstruct an attacker's plans from what the attacker leaves behind: logs, dropped files, a ransom note. Amazon Threat Intelligence got something closer to the attacker's own notes. Between October 3 and November 5, 2026, Amazon's MadPot decoy sensors posed as large language model (LLM) inference APIs, the endpoints applications call to get answers from an AI model. Several AI agents that connected to them sent along their own campaign memory. That is the running record an agent keeps of its goals, targets and results.

Amazon assesses with moderate confidence that the agents were running on ARTEX, an open-source "agentic" AI penetration-testing framework. It reports two separate campaigns. One was a broad hunt for stolen AI provider API keys. The other was a narrower and stranger effort to get past identity checks at two cryptocurrency exchanges with real-time deepfakes. Amazon also rates its view of who was behind the campaigns at moderate confidence.

How the decoys caught the agents

MadPot is Amazon's globally distributed network of decoy threat sensors. These are fake services set up to attract and record malicious traffic. For this research the sensors presented themselves as LLM-compatible chat interfaces. Agent frameworks send their planning context to whatever model endpoint they are pointed at. When the agents sent requests to the decoys, those requests included the agents' memory of the campaign.

The Hacker News describes how ARTEX is built. A "planner" agent breaks a high-level goal into smaller tasks and hands them to "worker" agents, and each worker calls an LLM to carry out its piece. That design is why the decoys saw so much. Each call to the model includes enough context for the model to do its job, and in this case that context was the attackers' own records.

Amazon is clear about what it does not know. It did not watch the activity happen live, and it says it cannot verify the outcomes the agents reported. When the agents' memory says a target was compromised or a key was harvested, treat that as an unverified claim made by the attacker's tooling.

Campaign one: AI keys used to find more AI keys

The larger campaign went after AI model credentials, meaning the API keys organizations use to pay for and reach commercial AI models. The agents targeted self-hosted AI gateways, orchestration tools and chat front-ends exposed to the internet. Amazon lists One API, New API, Veloera, Sub2API, LiteLLM, Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama.

Many of these apps exist to hold provider keys and route traffic to model APIs on behalf of a team. That makes a single exposed instance worth a lot. According to Amazon, the agents' playbook included:

  • LiteLLM's default master key. These are deployments where the administrative key was never changed from its default.
  • CVE-2025-3248 in Langflow.
  • CVE-2026-28797 in RAGFlow.
  • A header-based authentication bypass in the One API family, tied to an isCliRequest header.
  • Server-side request forgery (SSRF) through URL-fetch features to reach cloud metadata services. SSRF tricks a server into making requests for the attacker. Pointed at a cloud instance's metadata endpoint, it can expose the instance's own cloud credentials.

The agents also sorted what they found. Amazon says one operation graded every credential: L1 for a lead, L2 for a confirmed working API call, and L3 for quantified impact such as readable data or spendable balance. Keys that had been publicly exposed for more than 30 days were treated as low-value inventory, which suggests the operators cared most about fresh keys that had not yet been noticed and revoked.

The campaign was also designed to pay for itself. Amazon says the agents' records show stolen keys being fed into the framework's own model gateway, where they powered the agents searching for more keys. In practice, a victim's AI bill could end up funding the next round of attacks against other people's AI servers.

Campaign two: deepfakes at the KYC checkpoint

The second campaign targeted know-your-customer (KYC) verification at two crypto exchanges. KYC is the identity check where a new user uploads an ID and usually completes a live selfie or video step. According to the agents' records, they worked on two fronts: forged ID document photos and real-time face-swap deepfake video for the live check.

The records show tuning work. Amazon reports that a separate agent session built a real-time deepfake video pipeline on a single consumer-grade GPU and reached 11.5 frames per second. On the liveness check, the agents' best recorded attempt completed the full movement sequence with the face within 7 pixels of the target.

The outcome favored the defenders. In every attempt the agents recorded, the identity vendor's server-side verdict rejected the deepfake. A fake that looked close enough on the attacker's screen still failed the check that counted. The Hacker News notes that Amazon did not say whether the forged ID photos passed document checks.

Who was behind it

Amazon assesses with moderate confidence that the campaigns were run by disparate China-nexus threat groups, not one coordinated actor. It cites simplified Chinese in the prompts, China time-zone configurations, targeting consistent with Chinese interests and tooling predominantly used by China-nexus groups. Amazon has not named the groups. This is Amazon's assessment, not a confirmed attribution.

Amazon says it assessed activity referencing potentially affected AWS customers and notified them as appropriate.

A wider wave of AI attack tools

ARTEX is not a single tool sitting in one repository anymore. In a separate post, Cloudflare says ARTEX's creator, who uses the handle Autumn-27, took the project down on October 8. By then it had spread. Cloudflare counts more than 12,000 repositories and several hundred unique forks containing lineage from the original repository.

Cloudflare also gives a sense of scale for the whole category. In the last three days before its post, it saw more than 50 framework families attempt to scan its customers globally, and the number of IP addresses running such frameworks has grown into the thousands. It says daily requests from AI pentesting and exploitation tools have grown 30-fold since October 1. Some of these tools have begun chaining stolen customer credentials in more complex attacks, such as lateral movement.

The Hacker News reports one case from Cloudflare's incident responders. ARTEX gained access to a compromised account-level API token, previously used for CI/CD, that had no IP restrictions or rate limiting, and the activity that followed included 400 database dump requests in an hour. Cloudflare has not said whether those requests returned any data. The example still shows how quickly an automated agent can make use of one overly permissive token.

Penetration-testing frameworks are dual-use. Red teams build them for authorized testing, and criminals adopt them for the same reason defenders do: they automate tedious reconnaissance and exploitation. What Amazon's findings add is evidence that these agents are being used in sustained, financially motivated operations, with fixed targeting rules, credential grading and stolen resources fed back into the operation. Amazon also stresses that none of the techniques were new: the agents succeeded where basic hygiene was missing.

What It Means for Defenders

Who is affected: anyone running a self-hosted AI gateway, LLM proxy, agent builder or chat front-end that can be reached from the internet. That especially means LiteLLM, One API and its forks (New API, Veloera, Sub2API), Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama. These are often set up quickly by a developer or data team for a pilot and then forgotten, so start by finding out whether you have any.

  • Change default admin keys. If you run LiteLLM, confirm the master key is not the default. Do the same for every AI tool that ships with a built-in admin credential, and do not leave a new install unfinished where an outsider could claim it.
  • Take management endpoints off the public internet. Admin panels and management APIs for these tools should sit behind a VPN, an identity-aware proxy or an IP allowlist, not be open to anyone who scans for them.
  • Patch the named CVEs. Check Langflow deployments against CVE-2025-3248 and RAGFlow deployments against CVE-2026-28797, and move to the vendor's fixed release. Use each project's own advisory for fixed versions. For One API-family gateways, review whether the header-based authentication bypass Amazon describes affects your build, and update.
  • Rotate and scope provider keys. If any of these apps was exposed, assume the AI provider keys stored in it are compromised and rotate them. Give replacement keys the narrowest scope and spending limits your provider allows, and watch usage for unexpected spend. Amazon advises assuming exposed secrets will be found quickly.
  • Limit what the apps can reach. The SSRF-to-metadata technique only pays off if the AI app can reach the cloud metadata service and the credentials behind it are broadly privileged. Review outbound access and the permissions attached to the hosts running these tools.
  • Restrict CI/CD tokens by IP and rate-limit them. The Cloudflare case involved a token with neither control. Bind pipeline and API tokens to known source addresses, add rate limits, and alert on bulk export or dump-style requests.
  • Exchanges and fintechs: make sure the decision on identity verification is made server-side, as the vendor in Amazon's report did, and not by anything the client can influence. In the agents' own records, that server-side check rejected the deepfake every time.

For CMMC-bound contractors and small businesses, the rule is the same one that applies to any other server. An AI tool that holds credentials is a system holding credentials, and it belongs in your asset inventory, access-control and patch processes.

Sources