Hearth Guard

Hearth Guard is a model-independent security layer for tool-using AI agents. Agents triaging bug reports can be hijacked by prompt injections hidden in text, screenshots, or PDFs, causing them to leak secrets via tool calls. Hearth Guard solves this with a two-gate middleware and a OneMind supervisor that writes a hash-chained audit log. Gate 1 intercepts inputs, normalizing text, decoding blobs, running OCR, and scoring with Llama Prompt Guard 2 before the agent runs. Gate 2 is a safety net, using taint tracking across nine encodings and an egress filter to inspect the agent's tool calls. In our benchmark, the unprotected Muse Spark 1.3 agent leaked the canary token in 10.8% of attacks. With Hearth Guard, leaks dropped to 0% (0 of 83 cases) while maintaining 96% task completion. Limitation: The evaluation relies on synthetic attack corpora rather than real-world organic data.