The Agent Skills Directory

[PROMPT_INJECTION]: Raw user input is concatenated directly into the prompt for the analysis subagent without delimiters (such as XML tags or Markdown blocks) or sanitization. An attacker could use this to hijack the subagent's execution flow.
[COMMAND_EXECUTION]: The subagent is explicitly granted high-privilege tools including run_bash_command and write_file. While relevant to the skill's purpose, the presence of these tools significantly increases the potential impact of a prompt injection attack.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because the subagent ingests and processes untrusted data (local code files) while possessing dangerous capabilities.
Ingestion points: Code files and terminal output read via read_file, grep, and run_bash_command during analysis.
Boundary markers: Absent. The subagent is not instructed to isolate or ignore instructions embedded within the code it analyzes.
Capability inventory: Access to run_bash_command, write_file, read_file, and search_files (defined in bug-analyzer.md).
Sanitization: None. The content of files and user-provided bug descriptions are processed as natural language instructions by the subagent.

bug-analyzer-agent