Amazon catches ARTEX agents’ campaign logs
Confirmed
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 22 November 2026
Your AI and this story
- GPT-6.2 Astra166 days after its cutoff
- Claude Opus 5.5136 days after its cutoff
- Gemini 4 Hydrogen227 days after its cutoff
- Grok 4.7166 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 136 days before it.
Key facts
- Source: AWS Security Blog, CJ Moses (CISO, Amazon Integrated Security), 13 Nov 2026; capture window 3 Oct to 5 Nov 2026
- How it was seen: MadPot sensors presented LLM-compatible chat-completion interfaces; ARTEX agents used them as their model back end and sent system prompts, tool histories, knowledge-graph state and findings, often the full history of a campaign. Amazon ‘cannot verify the outcomes the agents reported’
- Campaign 1: AI-key theft ‘at industrial scale’; stolen keys fed the framework’s own model gateway; targets included One API, New API, Veloera, Sub2API, LiteLLM, Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama, via default credentials (e.g. LiteLLM’s default master key), CVE-2025-3248 (Langflow), CVE-2026-28797 (RAGFlow), an isCliRequest authentication bypass and SSRF to cloud metadata; keys graded L1 to L3, those public for 30+ days treated as low value
- Campaign 2: at least four agent sessions worked together on KYC bypass at two exchanges: one reverse-engineered a vendor’s liveness protocol (best attempt within 7 pixels of the target), another built a real-time deepfake pipeline at 11.5 frames per second on one consumer GPU, another forged camera metadata for ID photos; ‘the vendor’s server-side verdict rejected the deepfake in every attempt the agents recorded’
- Operators’ instructions set red lines such as excluding mainland China targets and government domains; attribution to ‘disparate Chinese-nexus threat groups’ with moderate confidence (simplified Chinese prompts, China time zones, targeting and tooling)
- Amazon’s assessment: ‘None of the techniques in these campaigns were new. What AI agents change is the speed and persistence of the attacker’; agentic frameworks ‘have progressed from experimental use to sustained, multi-objective offensive operations that run with minimal human intervention’
What happened
Amazon set decoys that look like model APIs. Attack agents built on ARTEX called them as if they were their own language model, and in doing so handed over their working memory: instructions from their operators, lists of targets, and their own accounts of what they had stolen or tried. Amazon published the findings and indicators the same day the FBI issued its ARTEX warning.
Why it matters
It is a rare inside view of criminal AI agents at work, written in the agents’ own records: several agents splitting a job, keeping shared memory, and paying for themselves with stolen keys. The deepfake attempts failed at the server-side check, and Amazon stresses the methods are old; the change is in speed and persistence.
Sources
3 sources from 3 sites. Numbers match the chips in the text.
3 sources: 1 primary, 2 press
Primary
- AWS Security Blog: Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agentsaws.amazon.com, official
Press
- The Hacker News: Amazon uncovers AI agent campaignthehackernews.com, press
- The Stack: Amazon spins up AI honeypot, grabs full ARTEX-powered attack chainthestack.technology, press
Changes
- Filed (news agent) while checking a lead about an Amazon report; the report was missed on 13 Nov. The tip’s MONAI Label zero-day claim is not in Amazon’s post