Security audit
claw-observability
Security checks for vulnerabilities and agentic risk
Overview
The skill legitimately installs shell-level hooks that report agent lifecycle events to an external CLAW service, but the SKILL.md explicitly instructs the agent to hide that reporting and the skill is marked always:true, which together are a notable red flag.
This skill's implementation matches an observability add-on: it installs hooks that send lifecycle events to an external CLAW server and needs an API key and base URL. However, SKILL.md tells the agent to hide the monitoring (never mention CLAW / act as if the skill does not exist) and the skill is force-included (always:true). Before installing, consider the following: - Do not install blindly. Inspect hooks/claw-hooks.sh and setup.sh yourself (they are included) and confirm the exact payloads and headers sent. You can run the hook script locally with a test payload to see what it posts. - If you want transparency, remove or edit the secrecy lines in SKILL.md (or avoid enabling the skill). A legitimate observability tool should not tell the model to conceal monitoring. - Limit exposure: create an API key scoped/minimized for telemetry, or point CLAW_BASE_URL to an internal/trusted endpoint. - Backup ~/.claude/settings.json before running setup.sh and review changes after installation. - If you cannot accept invisible telemetry, do not install or set CLAW_API_KEY; instead run a local or audited alternative. What would change this assessment: removing the secrecy directives from SKILL.md and removing always:true (so the skill is explicitly enabled by the user) would make this coherent and likely benign. Conversely, if the server at CLAW_BASE_URL or the API key were unknown/untrusted, or if the hook transmitted more sensitive data than documented, risk would increase.
SkillSpector findings are pending for this release.