Researchers find 26 LLM API routers secretly injecting malicious code and stealing developer credentials

UC researchers audited 428 third-party LLM API routers — intermediaries that sit between AI coding agents and model providers like Anthropic and OpenAI — and found 26 actively malicious: 9 injecting code into tool calls, 17 exfiltrating AWS credentials, and 1 draining ETH from a test wallet. The routers exploit the fact that they terminate TLS and see all plaintext payloads, yet no provider enforces cryptographic integrity between client and upstream model. The paper, "Your Agent Is Mine," is the first systematic study of this supply chain attack surface and highlights a growing risk for developers using AI agents for smart contract and wallet work.

TLDR by @Benthic

More coverage

• Researchers discover malicious AI agent routers that can steal crypto (tradingview.com)
• UC researchers warn third-Party AI routers are stealing crypto and ... (crypto.news)
• Will AI Steal Your Bitcoin? New Research Reveals 26 Malicious ... (ccn.com)
• Study Finds 26 Third-Party AI Routers are Secretly Stealing ... (techinpakistan.pk)
• ALERT: Researchers discover 26 third-party AI LLM routers secretly ... (x.com)

Explore the topic

• LLM
• API