Skip to content

Latest commit

 

History

History
70 lines (49 loc) · 1.96 KB

File metadata and controls

70 lines (49 loc) · 1.96 KB

MVAR Execution Boundary Specification

Version: 0.1 Status: Draft Reference Implementation: MVAR runtime (github.com/mvar-security/mvar)

Overview

This specification defines the minimum execution boundary contract for AI agent runtimes. It is designed to be runtime-agnostic and implementable by any agent framework.

1. Provenance Model

Every action submitted to a privileged execution sink MUST include:

Field Values Description
source model, user, system Origin of the action request
trust_level trusted, untrusted Trust classification
trace_id string Unique trace identifier

2. Critical Sink Categories

The following sink categories are defined as critical:

Sink Examples
shell bash, exec, subprocess
filesystem read, write, delete
network HTTP, TCP, DNS
credentials API keys, tokens, secrets
process_spawn fork, spawn, daemon

3. Enforcement Invariant

The core invariant of this specification:

UNTRUSTED input MUST NOT invoke CRITICAL sinks without explicit policy approval.

This invariant is deterministic. It does not depend on:

  • Prompt classification
  • Intent detection
  • Model confidence scores
  • Uncertainty quantification signals

4. Conformance

A runtime is considered conformant with this specification if:

  1. All critical sink invocations are mediated by a policy decision
  2. Provenance fields are propagated across agent boundaries
  3. Policy decisions are logged with trace_id binding
  4. Enforcement is fail-closed on missing or invalid provenance

5. Reference Implementation

MVAR implements this specification in full.

Verification: pip install mvar-security bash scripts/repro-validation-pack.sh

6. Versioning

This specification follows semantic versioning. Breaking changes increment the major version. Additive changes increment the minor version.

Changelog

  • v0.1 — Initial draft. Provenance model, sink categories, enforcement invariant.