Paper: 26 malicious LLM routers inject tool calls, steal keys, drain $500k
petrusenko_max · x · 2026-09-22
An arXiv paper, "Your Agent Is Mine," presents the first systematic study of malicious LLM API routers. Agents route tool calls through third-party proxies with full plaintext access, with no cryptographic integrity between client and upstream model.
- Researchers bought 28 paid routers (Taobao, Xianyu, Shopify) and collected 400 free ones
- 1 paid and 8 free routers actively injected malicious tool calls; 2 used adaptive evasion triggers
- 17 touched researcher-planted AWS canary credentials; 1 drained ETH from a planted private key; one case reportedly drained a client's $500k wallet
- Poisoning experiments: a leaked OpenAI key generated 100M GPT-5.4 tokens; weakly configured decoys yielded 2B billed tokens and 99 credentials across 440 Codex sessions, with 400 hosts potentially taken over within hours
A serious wake-up call for developers using third-party API routers or relays.
More from coding & agent
- Anthropic's Swiss cheese model explains why passing evals isn't enough for agents — hugobowne · 2026-09-22
- OpenAI's artists are now all using Codex in their workflow — andrew_n_carr · 2026-09-22
- TinyTorch: PyTorch's free curriculum to build an ML framework from scratch in 20 modules — PyTorch · 2026-09-22
- Adaptive reasoning for Codex tweaks effort mid-CoT based on task difficulty — daniel_mac8 · 2026-09-22
- End-to-end video editing and publishing pipeline built with Codex — brandon_galang · 2026-09-22
- Tip: use Codex to fit Blender parts with color-coded multi-angle screenshots — majidmanzarpour · 2026-09-22