The Agent Skills Directory

[PROMPT_INJECTION]: Instructions attempt to override standard agent behavior by enforcing 'AUTONOMOUS MODE' and explicitly commanding 'Do NOT ask questions', which is a pattern used to bypass user confirmation and safety filters.
[COMMAND_EXECUTION]: The skill executes a wide variety of shell commands across multiple languages and frameworks (e.g., npx, pytest, go test, cargo test, mvn, gradle, rspec) to run both existing and dynamically generated tests.
[EXTERNAL_DOWNLOADS]: The skill is instructed to automatically install test frameworks (e.g., Vitest, Jest, pytest) if they are not detected, which involves downloading and executing software from external registries at runtime.
[DATA_EXFILTRATION]: The skill performs comprehensive project-wide scans, specifically targeting sensitive modules such as authentication, payment, and data validation logic. While no direct network send is identified, this wide-scale access to sensitive code is a prerequisite for potential exfiltration.
[DYNAMIC_EXECUTION]: The skill dynamically generates test scripts based on the logic found in existing application code and then executes those scripts. This 'generate-then-execute' pattern can be exploited if the source code contains malicious instructions that influence the generated tests.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted project data and arguments without sanitization or boundary markers. 1. Ingestion points: Project source files and $ARGUMENTS. 2. Boundary markers: None present. 3. Capability inventory: Package installation (npm, pip), test execution (npx, pytest, go test), and file system writes. 4. Sanitization: None identified.
[PERSISTENCE]: The skill writes 'telemetry' data to a persistent file path in the user home directory ('~/.claude/projects/skill-telemetry.md'), which allows for tracking or data storage across different projects and sessions.

unit-test