The Agent Skills Directory

[PROMPT_INJECTION]: The skill's execution contract includes the directive "do not invent approval gates from caution." This instruction explicitly tells the agent to disregard its own safety-oriented heuristics or procedural filters that might otherwise restrict its actions, which is a common pattern for attempting to bypass platform guardrails.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and execute "candidates" (external implementations or tools) and read "external semantics" from documentation, which introduces a surface where malicious instructions in that content could influence the agent.
Ingestion points: The skill reads repository instructions, candidate source code, task-specific inputs, and external documentation during the benchmark definition and grading phases.
Boundary markers: The skill requires the use of clean Git worktrees and temporary directories to isolate candidate execution environments and prevent cross-run contamination.
Capability inventory: The skill utilizes a shell runner and execution wrapper to run arbitrary candidate code and capture stdout, stderr, and exit statuses for analysis.
Sanitization: It mandates the use of an independent oracle and blind correctness grading to ensure outcomes are verified against expected results rather than relying on potentially malicious candidate self-reporting.

ln-45-benchmark-comparator