Amazon says AI agents targeted LiteLLM, Dify apps for API keys
Amazon says captured AI agent records show hunts for API keys in LiteLLM, Dify and similar apps; rotate exposed keys and change defaults.
- Published
- Fixed in
- Langflow 1.3.0 for CVE-2025-3248; RAGFlow advisory lists no patched version, version 0.25.0 reportedly renders templates in a sandbox
- Exploitation
- Reported but unverified. Amazon says it cannot verify the agents’ reported outcomes; CISA lists CVE-2025-3248 in its KEV catalog.
Live scores
CVE-2025-3248- EPSS
- 1.000
very likely to be attacked soon - VEDAS
- 0.983
strong exploit evidence - On CISA KEV
- Yes, since 2025-05-05
Thin horizontal line: the “likely exploited” threshold (EPSS 0.10, VEDAS 0.5). Shaded “no data” areas are periods when cve-scores did not publish. The vertical line is the day CISA added it to its exploited list.
VEDAS flagged this bug about 15 days before CISA confirmed attacks.
Live scores
CVE-2026-28797- EPSS
- 0.006
low predicted chance of attack - VEDAS
- 0.057
little exploit evidence - On CISA KEV
- Not listed
- Detection rule
- No public rule yet
What happened
According to The Hacker News, Amazon’s threat intelligence team says AI agents used by attackers have been hunting internet-exposed AI apps such as LiteLLM, One API and Dify to steal the AI provider keys stored inside. The claim rests on the agents’ own records, which Amazon captured with decoy servers. Amazon says it did not observe the activity as it happened and cannot verify the outcomes the agents reported. It has moderate confidence that the agents ran on ARTEX, an open-source AI penetration testing tool, according to the same report.
Per The Hacker News, the decoy network is called MadPot and received agent requests from October 3 to late last week. Amazon assessed with moderate confidence that the operators were separate China-linked groups, citing Simplified Chinese prompts and China time zone settings, among other evidence. The records describe agents finding exposed AI apps, extracting stored provider keys, checking that they work, and adding them to a shared pool. The report says the records cover hundreds of targets and that Amazon notified AWS customers who may have been affected.
The Hacker News also reports that Cloudflare’s incident responders saw ARTEX use a compromised account-level API token that had no IP restrictions or rate limiting. According to the report, Cloudflare has not said whether the 400 database dump requests that followed returned any data, and how the token was first stolen has not been disclosed. The report adds that the FBI said in a FLASH alert that a range of attackers have used ARTEX since at least August, and that the alert summarizes 11 earlier reports and describes no new incident. This advisory relies on The Hacker News for these claims. The FBI alert itself could not be read from the data provided.
Not confirmed: whether any specific victim lost keys, how much was stolen, or which flaw (if any) was used against a given target.
Am I affected?
You may be, if you run any of the products the agent records name and they are reachable from the internet. According to The Hacker News, named targets include:
- LLM gateways and proxies: One API, New API, Veloera, Sub2API and LiteLLM.
- AI app builders: Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama.
The same report says the playbook in the records lists these entry points:
- Default administrator credentials, such as LiteLLM’s default master key.
- Unfinished installs that an outsider could take over.
- Management and provider settings endpoints that need no login.
- An authentication bypass through the isCliRequest header in One API family systems. The report says Amazon’s team names no affected versions or fix.
- Server-side request forgery (SSRF) through URL-fetch features, to reach cloud metadata services.
- Secrets left in client-side JavaScript, source maps, debug routes and environment files.
Two named flaws have official records:
- Langflow, CVE-2025-3248: CISA’s catalog describes a missing authentication flaw in the /api/v1/validate/code endpoint that lets a remote, unauthenticated attacker execute arbitrary code. CISA says it added the flaw to its Known Exploited Vulnerabilities catalog on 2025-05-05 and marks it as known to be used in ransomware campaigns, per CISA. The Hacker News reports that versions before 1.3.0 are affected and 1.3.0 is the fixed version.
- RAGFlow, CVE-2026-28797: the project’s advisory describes server-side template injection in the Agent Text Processing and Message components, which lets any authenticated user run operating system commands on the server. A registered account is required, and the advisory says self-registration is enabled by default. It lists v0.24.0 as affected and no patched version, per GitHub: infiniflow/ragflow advisory. The Hacker News reports that NVD lists 0.24.0 and earlier, and that version 0.25.0 renders templates in the sandbox the advisory recommends.
Not confirmed: whether CVE-2026-28797 was used in this campaign. The Hacker News mentions a separate, earlier RAGFlow case study dated August 26, in which the publisher had low confidence about which flaw, if any, let attackers run code and did not link those attacks to ARTEX. The article’s wording does not make clear which company published that case study, and this advisory does not attribute it to Amazon.
What to do now
- Find AI gateways, app builders and inference endpoints that are reachable from the internet, and treat them like other production systems, as Amazon’s team recommends per The Hacker News.
- Change default administrator credentials, such as LiteLLM’s default master key, and do not leave a new install unfinished where an outsider could take it over.
- Update Langflow to 1.3.0 or later. For RAGFlow, the advisory lists no patched version, so review version 0.25.0 and consider whether open self-registration (on by default, per the advisory) is needed.
- Search configuration files, environment variables, deployment scripts, CI/CD pipelines and databases for embedded secrets, and rotate exposed credentials quickly. The Hacker News attributes this advice to the FBI.
- Watch for misuse through unexpected AI spending or unusual usage, as Amazon’s team recommends.
- If you use Cloudflare, The Hacker News reports that API tokens can carry an expiry date and a client IP restriction (both optional settings). It also suggests looking for Worker scripts named artex-, artex-pgrecon-, artex-pgw-* or artex-do-*.
- Check logs against the indicators the report mentions: the 11 IP addresses published with the decoy findings, and the more than 2,500 addresses that researchers at Cisco Talos link to ARTEX. The report lists the researchers’ detection rules as Snort 3 rules 7:76, 7:77 and 7:78 and Snort 2 rules 1:67494, 1:67495 and 1:67496. Rules 7:78 and 1:67496 are off by default.