Overview
This series explains how SkillSpector turns an untrusted AI-agent skill into a bounded pre-install trust decision: it constrains acquisition, preserves raw and classified evidence, combines deterministic and optional semantic analysis, examines metadata, permissions, hooks, dependencies, and coverage, then maps findings to policy. A low score is posture evidence, not proof that every runtime path or opaque artifact is safe.
Learning outcomes
- SkillSpector constrains source acquisition and records whether evidence is complete, partial, opaque, or failed.
- Deterministic checks show patterns, code paths, data flows, dependency identity, signatures, and concealment; each answers a narrower question.
- Optional LLM analysis adds context but depends on provider availability, bounded batches, and an explicit egress policy.
- A safe-to-install decision is conditional on evidence, coverage, source relationships, and caller policy—not a guarantee of harmless runtime behavior.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
What does SkillSpector actually assess before installation?
SkillSpector assesses how a skill is acquired, what its bounded evidence contains, what code and data flows can do, how metadata and hooks shape agent behavior, and whether the resulting checks are complete. It does not authenticate the author or prove every runtime path is harmless.
Can a clean SkillSpector scan prove that an AI skill is safe?
No; a clean SkillSpector scan means no active finding was retained in the evidence that completed the applicable checks. It does not prove that opaque, uninspected, future, or runtime behavior is harmless.
What is the role of LLM analysis in SkillSpector?
LLM analysis adds bounded context about intent, paraphrased attacks, and purpose-versus-behavior mismatch; it does not replace deterministic findings or make an incomplete scan complete. Enabled mode sends analyzer-eligible content to the configured provider, so egress and provider availability affect the result.
How should a SkillSpector result guide an installation decision?
Use the recommendation with completeness, highest issue severity, source context, baselines, and local policy: allow only when the required evidence is complete and assumptions hold, warn when review is needed, and isolate or reject when blocking behavior or unacceptable unknowns remain.