Skillv1.0.0

ClawScan security

Eridian · ClawHub's context-aware review of the artifact, metadata, and declared behavior.

Scanner verdict

SuspiciousFeb 13, 2026, 3:32 PM

Verdict: suspicious
Confidence: medium
Model: gpt-5-mini
Summary: The skill's instructions broadly match a runtime hardening purpose, but naming inconsistencies, an embedded prompt-injection pattern, and the unknown origin make the package suspicious and worth manual inspection before use.
Guidance: This instruction-only skill largely contains appropriate guidance for hardening an agent, but several red flags mean you should not apply it blindly: (1) Verify the author and source — there is no homepage and the skill name in the registry (Eridian) does not match the SKILL.md (carapace/Pistolclaw). (2) Inspect the SKILL.md and reference files for any hidden or executable instructions (the file contains an 'ignore-previous-instructions' token flagged by the scanner). (3) Manually review and apply only the AGENTS.md edits you understand — do not let the skill autonomously modify AGENTS.md. (4) Put edits through a sandbox agent first and run the provided audit-template tests (prompt-injection, exfiltration, credential access). (5) If you plan to use parts of it, copy the patterns yourself from the references after sanitizing examples that include network endpoints or command snippets. If the publisher can explain the naming mismatches and remove the suspicious token, confidence would increase.
Findings: [ignore-previous-instructions] unexpected: This pattern was found in the SKILL.md. It is a common prompt-injection attempt to neutralize prior safeguards and is not expected for a legitimate security-hardening guide. Even if present as an example, its presence in the runtime instructions is suspicious and should be removed or clearly marked as inert example text by the author.

Review Dimensions

Purpose & Capability: noteThe description (runtime hardening, anti-exfiltration, approval flows) aligns with the SKILL.md content and reference docs. However metadata/name mismatches exist (registry entry 'Eridian' vs SKILL.md 'carapace' and occasional references to 'Pistolclaw'), and the skill has no verified source/homepage — these inconsistencies reduce trust and should be explained by the publisher.
Instruction Scope: concernSKILL.md is instruction-only and tells an agent to copy security patterns into AGENTS.md and to create a browser allowlist file. Those actions are within a hardening scope but they are direct instructions to modify agent configuration files. The document also contains an explicit prompt-injection example and a suspicious trigger string (see scan findings) that looks like an attempt to influence evaluation. Because this is the runtime instruction set, any hidden or contradictory text here can alter agent behavior — review carefully before applying.
Install Mechanism: okNo install spec and no code files — lowest technical risk from installation. The skill is instruction-only, so it won't automatically write binaries to disk, but it does instruct manual edits to AGENTS.md and creation of a JSON allowlist file.
Credentials: okThe skill requests no environment variables or credentials. It explicitly lists files that must not be read (openclaw.json, .env, *.key, etc.), which is consistent with its stated purpose.
Persistence & Privilege: notealways is false and the skill is user-invocable (normal). However the SKILL.md instructs modifying AGENTS.md to place rules 'near the top' so they are processed first; that is effectively a request to change agent configuration and therefore can alter persistent behavior. This is coherent with a hardening tool but warrants manual review of any AGENTS.md modifications before applying.