Change language to
0:00

Amazon’s records show AI agents hunting internet-exposed AI tools such as LiteLLM, Dify, and Langflow for stored provider keys.

The outcomes in the agents’ own records haven’t been verified. Amazon says it didn’t observe the activity as it happened and can’t confirm what the agents claimed to achieve. The records do show what the operators were after: keys that can run up bills on a developer’s AI accounts, or on the services built on them.

Amazon Threat Intelligence detailed the campaigns in an AWS Security Blog post by CJ Moses, CISO of Amazon Integrated Security. Its MadPot network of decoy sensors captured requests from AI agents between 3 October and 5 November 2026. The requests carried the agents’ operating memory from campaigns run on ARTEX, an open-source agentic AI penetration testing framework. Amazon assesses with moderate confidence that ARTEX was used, and that the operators are separate Chinese-nexus threat groups. The Hacker News followed with extra context from other companies. It says the MadPot decoys posed as popular LLM services, and that the records cover hundreds of targets.

Subscribe to our Newsletters for more Tech Stories

How the key hunt works

The most prominent campaign was organised as a self-sustaining operation in which stolen keys paid for the LLM inference that powered further credential hunting. According to Amazon’s records, the agents claim to have found exposed AI applications, extracted the upstream provider keys they held, validated them, and stored them in a shared pool. They favoured applications that aggregate keys, because one weak management interface can expose many downstream accounts.

Targets included LLM gateways and proxies such as One API, New API, Veloera, Sub2API, and LiteLLM. They also included application builders such as Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, and Ollama.

The playbook covered uninitialised installs that could be claimed, anonymous management endpoints, default admin credentials such as LiteLLM’s default master key, and SSRF to cloud metadata services. It also covered secrets left in client-side JavaScript, source maps, debug routes, and environment files. To find keys, the agents drew on internet scanning services, certificate transparency, GitHub, npm and PyPI tarballs, container image layers, paste sites, and Telegram channels.

FlawAffectsStatus
CVE-2025-3248LangflowFixed in 1.3.0
CVE-2026-28797RAGFlowFixed in 0.25.0, according to The Hacker News
isCliRequest header authentication bypassOne API familyNo fix stated

Amazon’s records also show a second campaign in which agents used real-time deepfake video and forged ID photos against KYC checks at two cryptocurrency exchanges. The vendor’s server rejected the deepfake in every attempt the agents recorded.

Amazon argues that none of the techniques was new. What AI agents change is the speed and persistence of the attacker. It says agents succeeded where basic hygiene was missing.

What others have seen

The Hacker News reports that Cloudflare saw ARTEX use a compromised account-level API token for CI/CD, with no IP restrictions or rate limiting. The activity included 400 database dump requests in an hour, and Cloudflare hasn’t said whether they returned data. According to The Hacker News, Cloudflare’s own findings say the token most likely came through a third-party cloud provider, though how it was first stolen hasn’t been disclosed. It also reports that the FBI issued a FLASH alert this week. The alert says a range of attackers have used ARTEX since at least August, and Microsoft reported attackers using such tools for GitHub secret hunting.

ARTEX’s developer made its GitHub repository unavailable on 8 October after reports of abuse, but forks and a commercial version, ARTEX PRO, remain available.

A checklist for developers

  • Change default admin credentials, such as LiteLLM’s default master key, and don’t leave a new install unfinished where an outsider could claim it.
  • Search configs, environment variables, CI/CD, and databases for embedded secrets, and rotate any exposed key quickly.
  • Treat AI apps, model gateways, and inference endpoints like any other production system, and assume exposed secrets will be found fast.
  • Watch for unexpected spend or unusual usage on your AI accounts.
  • Give Cloudflare API tokens an expiry and IP limits, and look for Worker scripts named artex-*, artex-pgrecon-*, artex-pgw-*, or artex-do-*.
  • Check logs against the 11 IP addresses in the AWS post. Cisco Talos links more than 2,500 addresses to ARTEX.
  • Update Langflow and RAGFlow to the fixed versions above.

For Gulf teams, the UAE Cybersecurity Council has already urged people to vet AI tools and their permissions. Tbreak’s earlier look at UAE businesses’ cyberattack exposure and identity gaps gives useful background on the weaknesses in Gulf firms.

Amazon says it notified AWS customers who may have been affected and shares its indicators so organisations running the affected tools can review their own environments.