Agent Security
PRYSM Engine
A deterministic runtime policy gate for AI agent tool calls. It sits between an agent and its tools and decides allow or deny from a versioned policy, never from prompt text.
TypeScriptPolicy engineHash-chained audit
Overview
PRYSM Engine is a deterministic, runtime policy gate for AI agent tool calls. It decides allow or deny from a versioned policy (never from prompt text) and records a privacy-aware, hash-chained audit trail.
What it does
- Enforces policy on tool calls routed through it: unknown tools, unauthorised tools, bad arguments and tainted-session external actions are blocked before the tool runs.
- In the default fail-closed mode, a call it cannot decide on does not go through.
What it does not do
- It does not detect or stop prompt injection itself. It limits what an injected agent can do.
- Provenance and session taint are asserted by the host and not verified by PRYSM.
- The audit chain is tamper-evident, not tamper-proof.
Status
MVP, not released. Milestones M0 to M5 are complete; M6, the final review and delivery report, is in progress. Measured results, with their limits, are documented in the repository.
Next project