research-entity ships only Markdown text — there is no executable code in the published artifact, no server, no auth flow, and no user data is collected or stored. The skill runs inside Claude Code, and the only outbound network calls it makes are the web searches and URL fetches the user's research request triggers.
This narrows the meaningful security surface to:
| Surface | Risk | Treatment |
|---|---|---|
| Factual content | A wrong claim about a real company could damage its reputation or lead a reader to a costly decision | Anti-hallucination contract + cross-validation + confidence labeling — report via the factual-correction issue template |
| Source URLs | A dead URL in a register/source file could redirect to a malicious or impersonating domain | Report via a regular issue with the replacement primary-source URL |
Dispatch logic in SKILL.md |
A crafted entity name or argument could route to the wrong template or skip verification steps | Report privately via Security Advisory (below) if exploitable |
| GitHub Actions workflows | CI reads repo files only | permissions: scoped to contents: read, no secrets, only actions/* actions |
Report privately via GitHub Security Advisory if you discover:
- A way to make the skill exfiltrate conversation data or credentials through a crafted research target
- A prompt-injection vector in a segment file that could override the anti-hallucination contract
- A supply-chain risk in the actions referenced in
.github/workflows/ - An impersonation issue (e.g. a fork republishing under a confusingly similar name)
Report publicly via regular issues for:
- Wrong facts in register/source files (use factual-correction)
- Dead or moved source URLs
- A dossier section that produced ungrounded output (that's a bug in the skill's guardrails)
- Acknowledgement target: 72 hours
- Coordinated disclosure preferred — please do not publish a working exploit before a patch lands
Only the latest state of main is actively maintained. There is no LTS branch.
Reporters of confirmed vulnerabilities are credited in CHANGELOG.md under "Security", unless they prefer anonymity.