The Agent Skills Directory

[COMMAND_EXECUTION]: In the analysis mode, the skill instructs the agent to "verify the repro command if present" using Bash. This pattern of executing commands derived directly from untrusted input text allows for arbitrary command injection if the input is maliciously crafted.\n- [REMOTE_CODE_EXECUTION]: The instruction to run shell commands provided within a bug analysis document constitutes a remote code execution vulnerability. An attacker could embed malicious payloads in the analysis target which the agent would then execute.\n- [PROMPT_INJECTION]: The skill processes untrusted external data (plans, code diffs, and analyses) without specifying boundary markers or sanitization logic. This exposes an attack surface for indirect prompt injection, where instructions hidden in the data could hijack the agent's behavior. Ingestion points include the $ARGUMENTS and data read via Read/Grep, with capabilities including file access and shell execution, and no explicit sanitization is provided.

critical