mathodology-agent-pipeline
Pass
Audited by Gen Agent Trust Hub on Sep 7, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to execute a local script
lint_run.pyto validate structured handoff blocks between agent phases. - Evidence: The text states "the lead lints every block with
lint_run.py handoff" in the Agent Handoff Format section ofSKILL.md. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process complex external data (mathematical contest problems) which drives the logic and code execution of the multi-agent pipeline.
- Ingestion points: Phase 0 in
SKILL.mdmaps clauses from the user prompt to determine contest types, scoring surfaces, and deliverables. - Boundary markers: The skill uses structured YAML blocks for handoffs (
handoff:block) to separate phase outputs, though it lacks explicit safety delimiters for the initial untrusted input. - Capability inventory: The workflow includes a 'Coder' role in Phase 4 that produces numerical computation, environment notes, and figures. It also involves PDF rendering in Phase 6 and automated linting via
lint_run.py. - Sanitization: The skill implements a robust internal audit system where every phase artifact must pass an "independent critic review" and a final "blind judge panel" to verify requirements and scoring risk before completion.
Audit Metadata