Runtime Attestation Probe

Helps validate that agent behavior at runtime matches the capabilities and constraints declared in its attestation. Detects divergence between what an agent...

This skill is conceptually reasonable for detecting conditional or environment-triggered misbehavior, but the runtime instructions are the security surface — and those instructions are currently broad and not tightly constrained. Before installing or running: 1) Review the full SKILL.md (ask the publisher for the complete runtime procedure) and demand explicit lists of files, paths, and endpoints the probe will access. 2) Only run the probe in an isolated sandbox or ephemeral VM that contains no real credentials (do not run it on production hosts). 3) Require that any probing of other skills be done with explicit, auditable consent and that logs be stored securely. 4) Prefer a signed/traceable implementation (code + release) rather than an instruction-only skill if you need repeated or automated probing. 5) Because the publisher and homepage are unknown, treat provenance as weak and exercise extra caution.

Type: OpenClaw Skill Name: runtime-attestation-probe Version: 1.0.0 The skill's stated purpose is to detect malicious runtime behavior (like conditional activation and data exfiltration) in *other* skills, not to perform such actions itself. The `SKILL.md` transparently describes the problem, the probe's functionality, and even its limitations and risks, emphasizing the need for sandboxed execution. While it requires `curl` and `python3`, these are justifiable for a tool designed to execute and monitor other processes for security analysis. The examples of credential harvesting and exfiltration are clearly presented as *outputs* of the probe, demonstrating what it *detects*, rather than instructions for the agent to execute.