The Agent Skills Directory

[SAFE]: No signs of prompt injection, obfuscation, or unauthorized data access were found. The skill's behavior is consistent with its stated purpose of investigating corporate records.
[INDIRECT_PROMPT_INJECTION]: The skill involves processing untrusted external data from the Companies House registry and user-supplied documents. This attack surface is well-managed through strict output structuring and instruction-based boundaries.
Ingestion points: SKILL.md and references/companies-house-investigation-model.md define the ingestion of live registry data and user-provided filings (PDF, CSV, HTML).
Boundary markers: Present. The skill mandates the use of specific tags such as registry_evidence, inference, and needs_verification to separate external data from agent logic.
Capability inventory: SKILL.md utilizes web search/browsing and file reading capabilities to retrieve registry data.
Sanitization: The instructions explicitly require the agent to cite sources for every fact and label risk signals as investigative leads rather than factual allegations.
[DATA_EXFILTRATION]: The skill focuses on retrieving public data from a well-known service (UK Companies House) and explicitly warns the agent to maintain the confidentiality of privileged or work-product material.

corporate-registry-investigation