Agent Skill Firewall / explore

Skill admission review

Reviewed September 7, 2026 · Independent educational resource

Record whether you have evidence for each control. Unknown answers remain unresolved. This tool does not inspect a skill or connect to an agent platform.

1. Publisher and exact revision identified · Admission condition

Can you tie the reviewed artifact to its publisher, immutable version and intended task?

2. Minimum necessary access established · Admission condition

Have data, writable paths, tools and network destinations been limited to a justified need?

3. Dependencies and installation reviewed

Have scripts, dependency sources and remotely loaded instructions been examined?

4. Identity and token scope validated

Are credentials resource-appropriate, scoped and protected from unnecessary exposure?

5. Execution isolation demonstrated

Have filesystem, subprocess and network restrictions been tested in this environment?

6. External effects governed

Do consequential writes, uploads and messages have appropriate policy or approval gates?

7. Logs and an accountable owner defined

Can an owner trace decisions without retaining unnecessary sensitive data?

8. Revocation and change review tested

Can you disable access and re-review a changed revision, scope or dependency?

How the result is determined

Every documented answer contributes one to the displayed count. The first two conditions override that count: an unidentified artifact or unjustified scope means admission conditions remain unresolved. With those conditions documented, any other missing or unknown answer requires further control review. Eight documented answers produce only a candidate for a controlled trial.

The rules are an original editorial framework, not an industry risk score. No answer proves that a control works. Read the proposed architecture and source guidance before adapting the review.