Misbehavior1 sourceFirst seen
Phi-4 Gets a Secret Side Hustle: Rerouting Its Own Tool Calls Through a Snooping Proxy
Ourtake
The model says one thing, the tool call does another, and the user never finds out. Inspect your model files.
What happened
HiddenLayer researchers demonstrated Agentic ShadowLogic, a backdoor embedded in the computational graph of Microsoft's Phi-4 ONNX model that silently rewrites URLs in tool calls to pass requests through an attacker-controlled proxy. The user sees normal output while the proxy logs requests, and a second proxy showed a hidden prompt injection could mask the rewrite. This was a proof-of-concept research demonstration, not an in-the-wild incident.