ProofAgent brings independent evaluation, runtime safeguards and governance to the AI agents organizations build, use and buy. Designed for local, private and organizational deployments. One platform, local-first, with the agent registry, policies, audit trail and records in ProofAgent Enterprise. The evaluation engine is open source, so anyone can inspect how the scoring works.
Every agent gets an identity, a risk tier aligned to the EU AI Act and a policy before it is tested. ProofAgent runs multi-turn adversarial conversations against the live agent, a multi-jury evaluation with three independent jurors score what it actually did, serious failures are re-tested, and the result is a PASS, REVIEW or BLOCK gate as a CI exit code, sealed in a Portable Evaluation Record on the open EIO-Agents standard that anyone can verify.
Guardrails powered by ProofAgent Nano, our breakthrough ultra-compact SLM for on-device AI protection, inspect what goes into an agent and what comes out: prompt injection, forged authority and hidden instructions on the way in; secrets, personal data and unsafe actions on the way out. Available today as a test bench in the ProofAgent Runner; a runtime gateway and SDK are planned.
A local proxy on the employee's own computer masks names, identifiers, cards, keys and secrets before a supported session leaves the device, and restores the real values in the reply. Supported for Claude Code, Codex and Cursor sessions started through the shield. Nothing on the machine is changed, and the masking never calls a cloud model.
Attach to a supported coding-agent session started through the Runner and see every intent on the device: files read, commands run, tools called, secrets touched, data leaving. Only risky intents are uploaded to the workspace. This is visibility, not enforcement; real-time blocking is planned.
The ProofAgent Runner runs evaluation, inspection, masking and session observation on the device. The enterprise workspace holds policies, the agent inventory, flagged session events and evidence. Records validate with public EIO-Agents tools, and release gates drop into CI. The method is described in papers published on arXiv.