wisdom-capture
Pass
Audited by Gen Agent Trust Hub on Oct 31, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources and scraping tools to synthesize content for the corpus.
- Ingestion points:
SKILL.mdStep 2 (URL enrichment) andURL-IMPORT.md(scraped metadata, spoken-word transcripts from Whisper, and captions from Playwright). - Boundary markers: There are no explicit instructions for the agent to use delimiters or ignore instructions embedded within the scraped text when performing analysis.
- Capability inventory: The skill possesses file system write access (
SKILL.mdStep 7), network access via web tools, and git operation capabilities (SKILL.mdStep 8, 9). - Sanitization: While basic formatting and whitespace cleaning are performed, there is no semantic sanitization of the content fetched from external URLs to prevent malicious prompts from influencing the agent's behavior during categorization or commit steps.
- [COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands that interpolate user-provided content.
- Evidence:
SKILL.mdreferences the execution ofwisdom_find_dup "$body" "$WISDOM_REPO"andgit commit -m "..."which includes a preview of the body content. - Risk: A maliciously crafted wisdom snippet could potentially exploit shell command construction if the agent or environment does not properly escape special characters in the
$bodyvariable during the duplication check or commit process.
Audit Metadata