The Agent Skills Directory

[PROMPT_INJECTION]: The skill interpolates the $ARGUMENTS variable directly into the instruction stream in SKILL.md without boundary markers or delimiters. This ingestion of untrusted data creates a surface for indirect prompt injection, potentially allowing a user-provided claim to override the agent's workflow.
[COMMAND_EXECUTION]: The workflow allows the agent to run shell commands or tests if a "single command or test can decide the claim." While intended for diagnostics, this provides an execution surface for potentially malicious commands if the investigation process is subverted by malicious input.
[DATA_EXFILTRATION]: The instructions direct the agent to read file paths, line numbers, code, and logs to gather evidence. A crafted input claim could lead the agent to access sensitive configuration files or codebase metadata as part of its verification process.

dialectic