Amazon honeypots capture AI agents autonomously stealing LLM provider API keys
ChuckDBrooks · x · 2026-11-17
Amazon's MadPot decoy sensor network captured records from AI agents describing a campaign to steal AI provider API keys between Oct 3 and Nov 5: the agents scanned internet-exposed AI apps, extracted upstream provider keys, verified them, and pooled them for reuse.
Key points:
- Targets focused on LLM gateways that centralize credentials from many providers, including One API, New API, Veloera, Sub2API, and LiteLLM; stolen keys reportedly paid for the models used to hunt more keys, a self-funding loop.
- Amazon assesses with moderate confidence the agents ran on ARTEX, an open-source AI pentesting framework, operated by separate China-nexus groups, citing Simplified Chinese prompts, China time zones, and targeting patterns.
- Amazon notes it did not observe the activity live and cannot verify the agents' claimed results.
A concrete security warning for anyone running LLM gateways or centralized key management.
More from AGI Musings
- Greentext Nails the AI Job Paradox: We Automated Pressing Buttons — _RodolfoOcampo · 2026-11-17
- Economist reworks his JPE 2000 knowledge-based hierarchy with AI as the top node — danielrock · 2026-11-17
- "Infohazard": researcher worries CoT scaling is quietly moulding his convictions — akbirthko · 2026-11-17
- Pedro Domingos: you need your own AI to interact with Claude and ChatGPT — pmddomingos · 2026-11-17
- Cyber incident base rates are rising, but AI panic is warping how we notice them — shakoistsLog · 2026-11-17
- Blogger claims AI will boost productivity 20x and urges kids to ditch 40-hour jobs — ZeroStateReflex · 2026-11-17