eagle3-triage
Pass
Audited by Gen Agent Trust Hub on Jul 6, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses local shell commands like ls, find, and tail to navigate experiment directories and extract log contents for diagnosis. It also uses the uv runner to execute local pipeline scripts such as launch.py for re-running tasks.
- [PROMPT_INJECTION]: The skill analyzes external data from pipeline logs, which constitutes a surface for indirect prompt injection if the training process or generated data contains adversarial instructions.
- Ingestion points: Pipeline log files (.out and .log) located in experiments/ subdirectories (SKILL.md).
- Boundary markers: Absent; log contents are read directly via tail without protective delimiters.
- Capability inventory: Execution of pipeline management scripts and shell-based log inspection (SKILL.md).
- Sanitization: No filtering or sanitization of log content is performed before the agent processes it.
Audit Metadata