adobe-document-review
Pass
Audited by Gen Agent Trust Hub on Sep 30, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted text from PDF documents, which creates a potential surface for indirect prompt injection.
- Ingestion points: Document text is extracted using the
pdf_to_markdowntool in SKILL.md. - Boundary markers: The skill contains specific guardrails instructing the agent to treat document content as untrusted data and to ignore any embedded directives or commands.
- Capability inventory: The agent can modify documents by adding highlights and comments through the
pdf_highlighttool. - Sanitization: The skill enforces verbatim string matching for highlights, which ensures that document content is handled as static text rather than executable instructions.
- [DATA_EXFILTRATION]: The skill includes logic to prevent unauthorized data access or exfiltration through malicious links.
- Verification: When document content is provided via a URL, the agent is instructed to verify that the link uses the
https://protocol and originates from an Adobe-issued asset host, specifically prohibiting access to local, private, or unauthorized IP addresses.
Audit Metadata