The Agent Skills Directory

[COMMAND_EXECUTION]: The skill frequently instructs the agent to execute shell commands for reconnaissance and auditing, including standard utilities like find and grep. It also executes bundled bash scripts such as scripts/scan_dead_code.sh, scripts/scan_patterns.sh, and scripts/scan_dependencies.sh to automate the hardening process.\n- [COMMAND_EXECUTION]: The agent is directed to invoke project-specific development tools including npx eslint, python -m pylint, go vet, and cargo clippy. These tools may execute code defined in the target project's configuration files or plugins.\n- [PROMPT_INJECTION]: The skill performs an audit of an entire codebase, which involves reading and processing untrusted data. It lacks specific instructions or boundary markers to isolate file content from the agent's control logic, creating a surface for indirect prompt injection where malicious comments or code strings in the audited files could influence the agent's auditing conclusions or subsequent actions.\n
Ingestion points: Reads all source, configuration, and documentation files within the user-provided codebase directory during Phase 1 (Pass 1-3).\n
Boundary markers: None provided to separate audited content from agent instructions.\n
Capability inventory: Shell command execution, execution of bundled bash scripts, and writing to the file system (AUDIT.md).\n
Sanitization: No sanitization or escaping is performed on the content read from the target codebase.\n- [REMOTE_CODE_EXECUTION]: The scripts/scan_patterns.sh script uses node -e to call require() on a package.json file within the audited directory. Because require() in Node.js can execute JavaScript code if a malicious .js file is placed to shadow the JSON or if the file itself contains executable exports, this represents a potential dynamic code execution risk when auditing untrusted directories.

iblai-deslop