Instrument A · Evidence Plane 1

SafeAI — Agent Authority Evidence

An open-source static analyser for understanding the authority an AI agent can potentially exercise.

Current model

Discover → Compare → Govern

01
Discover
What can the agent access or control?
tools, MCP servers, skills, destinations, data reachability
02
Compare
What changed from the previously approved state?
authority diff, access-mode transitions
03
Govern
Should the authority change be allowed?
ChangeGuard, governance controls, evidence contracts

Example

An authority diff

  agent: release-bot            principal: platform-team
  tool:  github.repo             mode: read
- tool:  github.repo             mode: read
+ tool:  github.repo             mode: write      ← access-mode transition
+ mcp:   deploy-server           mode: unknown    ← UNKNOWN AUTHORITY
  ─────────────────────────────────────────────────────
  decision: REVIEW   reason: write escalation + unknown authority

Illustrative example only — not output from a real assessment.

Scope

Concepts SafeAI works with

KYA
Agent Authority
ChangeGuard
capability discovery
authority diff
access-mode transitions
MCP analysis
tool analysis
governance controls
evidence contracts
provenance
unknown authority
IaC authority evidence

Unknown is an evidence state, not evidence of safety.

SafeAI is sometimes compared to “SonarQube for AI” as historical context. Its focus is narrower and deeper: the reachable authority of an agent, and how that authority changes.