Skip to content

nabinpkl/agent-action-classifier

Repository files navigation

agent-action-classifier

A policy-driven governance classifier for AI agent actions. Given an agent action and an organization policy, it decides allow / deny / escalate / flag, records which OWASP Agentic clause fired and why, and emits an audit-defensible decision log.

What it is

Built as the standard authorization architecture (XACML's P*Ps): a pure, environment-independent Policy Decision Point in Rust, driven from a Python host that holds the impure edges (enforcement adapter, context/approval source, audit sink). The decision point runs a layered cascade: a cheap deterministic rule engine settles the clear cases at sub-millisecond latency, and an LLM judge is consulted only for the ambiguous, semantic ones. Organization policy is supreme: an explicit org deny can never be overridden by a local user approval.

It is a learning project: the goal is to understand agent-action governance, rule-engine internals, and the audit trail by building the load-bearing parts by hand and measuring them, not to ship a product.

Status

v0 is scoped and accepted; implementation has not started. v0 proves one vertical slice end to end, the OWASP ASI05 (Unsafe Code Execution) clause over shell and file-write actions, and measures it.

Documentation

  • PRD.md — what we build and why (product intent).
  • SPEC.md — technical shape (in progress).
  • docs/adr/ — architecture decisions, one per file.
  • AGENTS.md — how to operate in this repo.

License

No license yet. All rights reserved while the project is experimental.

About

Experimental learning project: a policy-driven governance classifier for AI agent actions (OWASP Agentic Top 10), with a Rust policy-decision core. WIP, not for production.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors