The Hacker News
Amazon says AI agents targeted exposed apps to steal API keys
Part of the CVE-2025-3248 story · since 26 Mar
- Thu 26 Mar, 22:02 UTC. Security Affairs, first report. U.S. CISA adds a Langflow flaw to its Known Exploited Vulnerabilities catalog. Cites a CVE in CISA KEV.
- Fri 3 Jul, 11:32 UTC. SecurityWeek, its first report on the story. Agentic AI Used to Conduct Ransomware Attack via Langflow. Cites a CVE in CISA KEV.
- Fri 3 Jul, 12:32 UTC. Security Affairs. JADEPUFFER AI ransomware exploits Langflow flaw CVE-2025-3248. Cites a CVE in CISA KEV.
- Mon 6 Jul, 17:41 UTC. Dark Reading, its first report on the story. LLM Driven Ransomware Automates Attack via Langflow CVE-2025-3248. Cites a CVE in CISA KEV.
- Fri 10 Jul, 08:41 UTC. SecurityOnline, its first report on the story. JADEPUFFER Marks the First AI-Driven Agentic Ransomware Attack. Cites a CVE in CISA KEV.
- Wed 30 Sep, 15:01 UTC. Google Cloud, its first report on the story. Google Warns N Day Flaws Are Driving a Surge in Exploitation. Cites a CVE in CISA KEV.
- Mon 16 Nov, 08:30 UTC. The Hacker News, its first report on the story. Amazon says AI agents targeted exposed apps to steal API keys. Cites a CVE in CISA KEV. This article, 7th of 8.
- Mon 16 Nov, 11:00 UTC. Security Affairs. Amazon finds cybercriminals using AI agents to scale attacks. Cites a CVE in CISA KEV.
Amazon’s threat intelligence team says attackers operating AI agent systems have been hunting internet‑exposed AI apps such as LiteLLM, One API, and Dify to exfiltrate provider keys stored within. The decoy‑server campaign records, captured via AWS decoys, describe stolen keys being used to run the AI models that searched for more keys. The team cautions it cannot verify the agents’ outcomes but has moderate confidence that ARTEX, an open‑source AI pentesting tool, powered the operation.
The playbook outlines entry points, including LiteLLM’s default administrator credentials, unfinished installations, and known flaws in Langflow and RAGFlow. Separately, Cloudflare incident responders observed ARTEX gaining access to a compromised API token used for CI/CD, followed by a large number of database dump requests; it remains unclear if data were returned.
The records describe a two‑step workflow: an operator assigns targets and goals, a planner splits tasks for worker agents that run security tools, and a large language model guides each next move. The decoys, under the MadPot network, posed as popular LLM services and the attackers’ requests carried full campaign history. Amazon notes the actors were likely China‑linked groups, based on Simplified Chinese prompts and time‑zone cues.
Targets include LLM gateways and proxies (One API, New API, Veloera, Sub2API, LiteLLM) and AI app builders (Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, Ollama). The incident note lists unintended authentication bypasses in One API, SSRF via cloud‑metadata access, and secrets left in client code; it also highlights the ongoing responsibility to protect cloud tokens, with FBI FLASH alert cited as context.
Flaws named include Langflow (CVE-2025-3248; affected before 1.3.0), RAGFlow (CVE-2026-28797), and One API authentication bypass via isCliRequest, with further checks and mitigations listed.
How the story developed
8 articles from 6 outlets since 26 Mar. Picked up again on 16 Nov after 7 weeks quiet. This was the 7th report, 234 days after the first.
- Thu 26 Mar, 22:02 UTC. Security Affairs, first report. U.S. CISA adds a Langflow flaw to its Known Exploited Vulnerabilities catalog. Cites a CVE in CISA KEV.
- Fri 3 Jul, 11:32 UTC. SecurityWeek, its first report on the story. Agentic AI Used to Conduct Ransomware Attack via Langflow. Cites a CVE in CISA KEV.
- Fri 3 Jul, 12:32 UTC. Security Affairs. JADEPUFFER AI ransomware exploits Langflow flaw CVE-2025-3248. Cites a CVE in CISA KEV.
- Mon 6 Jul, 17:41 UTC. Dark Reading, its first report on the story. LLM Driven Ransomware Automates Attack via Langflow CVE-2025-3248. Cites a CVE in CISA KEV.
- Fri 10 Jul, 08:41 UTC. SecurityOnline, its first report on the story. JADEPUFFER Marks the First AI-Driven Agentic Ransomware Attack. Cites a CVE in CISA KEV.
- Wed 30 Sep, 15:01 UTC. Google Cloud, its first report on the story. Google Warns N Day Flaws Are Driving a Surge in Exploitation. Cites a CVE in CISA KEV.
- Mon 16 Nov, 08:30 UTC. The Hacker News, its first report on the story. Amazon says AI agents targeted exposed apps to steal API keys. Cites a CVE in CISA KEV. This article, 7th of 8.
- Mon 16 Nov, 11:00 UTC. Security Affairs. Amazon finds cybercriminals using AI agents to scale attacks. Cites a CVE in CISA KEV.
- Known exploited CVE
- No CVE
- Larger: an outlet’s first report
- Coverage over time
- Quiet time, folded
Show all 8 articles as a listHide the list
- Thu 26 MarDay 01
- U.S. CISA adds a Langflow flaw to its Known Exploited Vulnerabilities catalog22:02 · Security Affairs · First report · Exploited CVE · Known exploited CVE
Quiet for 14 weeks
Fri 3 JulDay 992- Agentic AI Used to Conduct Ransomware Attack via Langflow11:32 · SecurityWeek · Picked up again · Known exploited CVE
- JADEPUFFER AI ransomware exploits Langflow flaw CVE-2025-324812:32 · Security Affairs · Known exploited CVE
Quiet for 3 days
Mon 6 JulDay 1021- LLM Driven Ransomware Automates Attack via Langflow CVE-2025-324817:41 · Dark Reading · Picked up again · Known exploited CVE
Quiet for 4 days
Fri 10 JulDay 1061- JADEPUFFER Marks the First AI-Driven Agentic Ransomware Attack08:41 · SecurityOnline · Picked up again · Known exploited CVE
Quiet for 12 weeks
Wed 30 SepDay 1881- Google Warns N Day Flaws Are Driving a Surge in Exploitation15:01 · Google Cloud · Picked up again · Known exploited CVE
Quiet for 7 weeks
Mon 16 NovDay 2352- Amazon says AI agents targeted exposed apps to steal API keys08:30 · The Hacker News · Picked up again · Known exploited CVE · This article, 7th of 8
- Amazon finds cybercriminals using AI agents to scale attacks11:00 · Security Affairs · Known exploited CVE