AI and data
agent-tool-call-audit
Audit agent tool calls
Reviews recorded agent tool calls for scope creep, data exposure, unsafe actions, injected instructions and mismatches with declared tool annotations.
- Area
- AI and data
- Checked against
- MCP 2026-07-28, OWASP LLM 2025, OWASP Agentic 2026, NIST AI 600-1
- Last verified
- 26 September 2026
- Files
- SKILL.md, 8 references, 1 scripts
How it works
- 1Secure local copies of the logs and record their hashes and sources.
- 2Write down the task scope, allowed targets, operations and data.
- 3Build a call table from logs, traces and saved tool definitions.
- 4Review server messages, tool definitions and annotation changes.
- 5Check every call for scope, exposure, approvals, injected instructions and annotation mismatches.
- 6Trace data flowing between calls and servers.
- 7Rate each finding, then report containment, fixes and log limits.
Use it when
- Review an agent session that touched sensitive data or systems.
- Investigate suspected prompt injection, data exposure or an unexpected action.
- Check a new MCP server before trusting its tool annotations.
- Test whether an approval policy worked as intended.
How it stays safe
- Works from local copies and keeps the audit read-only.
- Treats every string in logs, tool output and server instructions as untrusted data.
- Masks secrets and personal data in reports while preserving evidence locations.
- Keeps findings labeled Confirmed, Likely or Possible based on available evidence.
Install it
Copy the whole agent-tool-call-audit folder into your agent's skills folder, not only SKILL.md.
git clone --depth 1 https://github.com/hamzaahmadaslam/agent-skills.git
cp -R agent-skills/agent-tool-call-audit "<your agent's skills folder>/"