devonthink
Pass
Audited by Gen Agent Trust Hub on Mar 15, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes
osascript -l JavaScriptcommands to perform automation tasks within the local DEVONthink application environment. - [EXTERNAL_DOWNLOADS]: The JXA API reference in
references/jxa-api.mdincludes functions for the agent to download remote content and data into the local database, such asapp.downloadURLandapp.createMarkdownFrom(url). - [PROMPT_INJECTION]: The skill's ability to ingest and process document content presents an indirect prompt injection surface.
- Ingestion points: Untrusted data can enter the agent context through bibliography JSON files read by
scripts/bib_lookup.pyand via theplainTextcontent of DEVONthink records accessed inSKILL.md. - Boundary markers: No specific delimiters or boundary markers are documented to isolate potentially malicious instructions embedded in documents from legitimate content.
- Capability inventory: The skill maintains capabilities for local command execution via
subprocessinscripts/bib_lookup.pyandosascriptinSKILL.md, alongside file system access and network operations via the DEVONthink API. - Sanitization: Document content is retrieved and used in AI summarization or chat workflows without sanitization or filtering of embedded instructions.
Audit Metadata