summer-school-finder

Warn

Audited by Gen Agent Trust Hub on Apr 11, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands that incorporate variables ({PROGRAM_URL} and {SAFE_NAME}) sourced from web searches and agent-generated identifiers. Interpolating these values directly into a bash command string (e.g., tinyfish agent run --url "{PROGRAM_URL}" ... > /tmp/ssf_{SAFE_NAME}.json &) without explicit sanitization or validation instructions creates a command injection vulnerability if the variables contain shell metacharacters such as double quotes, semicolons, or backticks.\n- [PROMPT_INJECTION]: The skill scrapes and processes untrusted content from multiple university program pages, creating an indirect injection surface.\n
  • Ingestion points: Data is fetched via the tinyfish tool, saved to files in /tmp, and then read back into the agent's context using cat in SKILL.md.\n
  • Boundary markers: While the tinyfish agent is given limits on its own navigation, the main agent processes the aggregated results without clear delimiters or instructions to ignore embedded instructions in the scraped data.\n
  • Capability inventory: The skill possesses shell access (including the ability to run the tinyfish tool and standard utilities) and file system read/write access to /tmp.\n
  • Sanitization: No sanitization, escaping, or schema validation is performed on the scraped content before it is presented to the agent for ranking and analysis.\n- [EXTERNAL_DOWNLOADS]: The skill performs outbound network requests to 7-8 different external university websites to scrape program details. While this is the intended functionality, it involves interacting with untrusted third-party domains.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Apr 11, 2026, 04:42 PM