runtime-attestation-probe_skill

Overview

This skill helps validate that an agent's runtime behavior matches the capabilities and constraints declared in its attestation. It detects conditional activation, undeclared resource access, unexpected data flows, and other behaviors that static checks miss. The probe produces a concise report with a verdict and an attestation drift score to guide remediation.

How this skill works

The probe executes the target skill in multiple controlled environments and collects traces of file, network, and system interactions. It compares observed behaviors against the provided attestation claims across capability boundaries, data flows, side effects, and conditional triggers. The output is a structured runtime attestation report highlighting violations, conditional activation patterns, and a drift score (0–100).

When to use it

• Before deploying a skill to sensitive or production environments
• When an attestation exists but you need runtime confirmation
• After a version update or untrusted contributor change
• When conditional or environment-based attacks are a concern
• During security reviews of third‑party or archived skills

Best practices

• Run probes in isolated sandboxes with no access to real credentials or production systems
• Test multiple environment permutations (env vars, file presence, execution counts) to reveal conditional logic
• Use representative execution samples per environment to reduce false negatives
• Correlate probe findings with publisher metadata and version history before taking action
• Treat high drift or undeclared outbound endpoints as immediate red flags and isolate the skill

Example use cases

• Validate a reporting skill that claims only local reads/writes to ensure no hidden exfiltration
• Scan archived or marketplace skills for environment‑triggered credential harvesting patterns
• Compare observed outputs of a skill against its attestation after a publisher update
• Triage a flagged skill by reproducing conditional behavior in a production‑like sandbox
• Integrate into a pre-deployment pipeline to block skills with capability drift