Capafy logo Capafy
开始使用
Agent Governance & Evidence Auditor

Agent Governance & Evidence Auditor

by: Wy and Adams
Audit AI agent workspaces for authority, secrets, data flow, memory, supply-chain, and evidence risks—locally and without executing target code.
#工程#分析#咨询
评分
需要更多评价
订单数
0
如何使用
下载

Agent Governance & Evidence Auditor

Audit AI agent workspaces without executing the code being reviewed.

Agent Governance & Evidence Auditor is a deterministic, read-only auditing tool for AI agent projects, Skills, MCP servers, tool manifests, model gateways, persistent memory, automation configurations, and related source files.

It runs locally on your own hardware and produces an evidence-backed governance and security readiness report.

What It Reviews

The auditor checks for risks and control gaps related to:

  • Agent authority and excessive permissions
  • Wildcard tool access
  • Hardcoded credentials and secret exposure
  • Unsafe shell and dynamic code execution
  • Undeclared external data destinations
  • Prompt-injection patterns
  • Persistent-memory and knowledge-promotion risks
  • Dependency and supply-chain hygiene
  • MCP and Skill inventory
  • Evidence integrity and audit readiness

Read-Only Observation Mode

The auditor is designed to inspect an exported or local workspace without changing it.

It does not:

  • Execute target code
  • Import target modules
  • Run discovered commands
  • Install target dependencies
  • Start MCP servers or agent tools
  • Upload workspace content
  • Apply automatic remediations

The selected workspace is treated as untrusted input throughout the audit.

Output

Each audit can generate:

  • Interactive offline HTML report
  • Printable risk and compliance readiness PDF
  • Machine-readable JSON findings
  • Workspace inventory JSON
  • CSV risk register
  • Redacted NDJSON evidence
  • SHA-256 integrity manifest

Findings include:

  • Severity
  • Confidence
  • Control ID
  • File and line reference
  • Redacted evidence
  • Framework mappings
  • Remediation guidance
  • Verification steps
  • False-positive conditions

Local Review UI

The generated HTML report includes:

  • Risk score
  • Finding totals
  • Severity filters
  • Control mappings
  • Evidence locations
  • Remediation guidance

The optional local viewer binds to 127.0.0.1 only.

No cloud dashboard or external account is required.

Example Use Cases

Before Installing an Agent or MCP Server

Review permissions, network destinations, install scripts, secret handling, and unsafe execution patterns before connecting an unfamiliar agent component.

Before Publishing an Agent Skill

Check a Skill workspace for credentials, risky authority grants, undeclared data flows, dependency issues, and evidence gaps before commercial distribution.

Internal AI Governance Review

Generate a structured risk register and evidence package for engineering, security, governance, and compliance-readiness discussions.

Persistent Memory Review

Identify long-lived memory writes or knowledge promotion behavior that may lack provenance, validation, confidence, expiry, or human review.

Quick Start

Run an audit:

python scripts/aga.py scan \
  --workspace /path/to/agent-workspace \
  --output /path/to/audit-report \
  --pdf
Verify the integrity of the generated report:
python scripts/aga.py verify \
  --bundle /path/to/audit-report
Start the optional local report viewer:
python scripts/aga.py serve \
  --bundle /path/to/audit-report \
  --host 127.0.0.1
Then open:
http://127.0.0.1:7841/report.html
Requirements
Python 3.10 or newer
Windows or Linux
No Python package installation required
No external API key required
No internet connection required
Validated environments include:
Windows with Python 3.12
CachyOS Linux with Python 3.14
Ubuntu-family Linux with Python 3.12
Privacy and Data Handling
The auditor reads only the workspace selected by the user and writes only to the selected report directory.
It does not include telemetry, analytics, cloud storage, or external data transmission.
Potential secrets found during an audit are redacted before evidence is written. File names, line numbers, hashes, and redacted snippets may still contain sensitive security context, so generated reports should be protected as internal security evidence.
Framework Readiness Mappings
Findings may include readiness mappings to public guidance such as:
NIST AI Risk Management Framework
NIST Secure Software Development Framework
OWASP Agentic AI security guidance
OWASP Agentic Skills security guidance
These mappings help organize review work. They do not constitute certification.
Important Limitations
This product is a deterministic static auditor.
It does not prove that an agent is safe, secure, compliant, or free from vulnerabilities. It cannot observe runtime behavior that is absent from the supplied workspace.
Results require human validation, particularly for Critical and High findings.
This product is not:
Legal advice
A penetration-testing tool
An ISO certification
An IEC certification
Proof of regulatory compliance
A replacement for qualified security or governance review
Support
Publisher: WY & Adams
Support email: [email protected]
When requesting support, do not send passwords, API keys, private keys, customer data, or production credentials. Provide the minimum redacted information needed to reproduce the issue.

View more