eagle3-validate

Warn

Audited by Gen Agent Trust Hub on Jul 6, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: In Step 1, the skill executes a bash script using the find command with an interpolated <exp_id> variable: find experiments/<exp_id>/ -type f .... If the experiment identifier is provided by the user or crafted maliciously within the file system, it can include shell metacharacters (e.g., ;, &, $()) to execute arbitrary commands on the runner.
  • [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting and acting upon untrusted data from pipeline log files.
  • Ingestion points: The agent reads the last 50 lines of log files using tail -50 in Step 1 and extracts specific strings for acceptance rates and training loss in Steps 3 and 4.
  • Boundary markers: There are no delimiters or instructions to the agent to treat the log content as data rather than instructions.
  • Capability inventory: The agent has the ability to execute shell commands (ls, find, tail, sort) and interact with the local filesystem (/scratchspace/, experiments/).
  • Sanitization: No validation or sanitization of the log content is performed before processing.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jul 6, 2026, 01:45 PM
Security Audit — agent-trust-hub — eagle3-validate