ablation-planner

Pass

Audited by Gen Agent Trust Hub on Jul 18, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill ingests data from local project documents such as research contracts and experiment logs to inform its planning process, which introduces a surface for indirect prompt injection.
  • Ingestion points: Reads project files including docs/research_contract.md and EXPERIMENT_LOG.md in Step 1.
  • Boundary markers: While the Codex prompt is structured with clear labels, it lacks explicit safety instructions to disregard commands found within the ingested project data.
  • Capability inventory: The skill has the ability to write files and execute shell commands via the Bash, Write, and Edit tools.
  • Sanitization: Content from project files is interpolated into prompts without escaping or sanitization.
  • [COMMAND_EXECUTION]: The workflow involves generating and running shell scripts to execute experimental ablations. This is an expected and necessary function for the skill's stated purpose of automating research tasks.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 18, 2026, 04:51 PM
Security Audit — agent-trust-hub — ablation-planner