jina-ai
Fail
Audited by Gen Agent Trust Hub on Aug 10, 2026
Risk Level: HIGHDATA_EXFILTRATIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The
scripts/dedup_images.pyscript included in the skill contains logic that reads arbitrary local files and transmits their contents to an external server. - Evidence: The
to_imagefunction inscripts/dedup_images.pyaccepts a path as an argument and usesPath.read_bytes()to read the entire file if it exists locally. - Evidence: The binary content is then base64-encoded and sent via a POST request to
https://api.jina.ai/v1/embeddings. - The script lacks file-type validation, meaning it could be coerced into reading and exfiltrating sensitive files (e.g., SSH keys, environment variables) if they are passed as arguments instead of images.
- [COMMAND_EXECUTION]: The skill's instructions recommend the use of shell pipes and
xargsfor parallel processing of queries, which can introduce command injection risks if inputs are maliciously crafted. - Evidence: Examples such as
jina expand "LLM" | xargs -P 5 -I {} jina search "{}"rely on the shell's expansion behavior and could be manipulated by unexpected output from subcommands. - [EXTERNAL_DOWNLOADS]: The skill installs external software and fetches academic data from well-known public sources.
- The skill installs the
jina-clitool using theuvtool manager. - It fetches PDF files and research papers from
arxiv.organd search results fromssrn.com. - [PROMPT_INJECTION]: The skill creates an attack surface for indirect prompt injection by ingesting and processing data from untrusted external sources.
- Ingestion points: Web page content extracted via
jina read, search snippets fromjina search, and academic paper text from PDF extractions. - Boundary markers: The instructions do not specify the use of delimiters or isolation protocols when the agent processes the ingested data.
- Capability inventory: The agent has access to local file system reading (via
scripts/dedup_images.py), network transmission capabilities, and shell command execution. - Sanitization: No explicit content sanitization or filtering logic is provided for handling untrusted external data.
Recommendations
- AI detected serious security threats
Audit Metadata