the-competitor-dossier
Pass
Audited by Gen Agent Trust Hub on Aug 14, 2026
Risk Level: SAFEPROMPT_INJECTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill retrieves and processes untrusted external content from the web, which represents a surface for indirect prompt injection attacks where malicious instructions are hidden in data.
- Ingestion points: The skill uses
WebFetchto read competitor homepages and pricing pages, andWebSearchto ingest reviews and news. It also processes user-provided inputs such as "pasted exports" or tool stack descriptions. - Boundary markers: The instructions in
references/agent-security.mdprovide natural language boundaries, directing the agent to treat all retrieved content strictly as data, use quoting for reporting, and explicitly ignore embedded instructions. - Capability inventory: The skill utilizes
WebFetchandWebSearch. It does not have shell execution, file-writing, or network-post capabilities enabled in the frontmatter configuration. - Sanitization: There are explicit requirements to report injection attempts as findings rather than obeying them, and to redact any detected credentials from output.
- [DATA_EXPOSURE_AND_EXFILTRATION]: The skill is designed to handle potentially sensitive inputs such as CRM exports and call transcripts, which may contain hardcoded credentials.
- Evidence: Rule 2 in
references/agent-security.mdestablishes a protocol for the agent to detect, flag, and avoid echoing API keys, session tokens, or passwords found in the data, recommending rotation instead of reproduction. - [EXTERNAL_DOWNLOADS]: The skill uses platform-provided tools to fetch content from external URLs as part of its core functionality.
- Evidence: The YAML frontmatter permits the use of
WebFetchandWebSearchto download and process content from competitor websites and third-party review platforms.
Audit Metadata