vibe-coding-polisher

Pass

Audited by Gen Agent Trust Hub on Aug 11, 2026

Risk Level: SAFE
Full Analysis
  • [PROMPT_INJECTION]: The skill includes an adversarial test fixture in 'evals/fixtures/execution-safety/prompt-injection-source/README.md' containing a typical jailbreak attempt ('IGNORE ALL PREVIOUS SAFETY INSTRUCTIONS'). This is explicitly documented as data used for evaluating the agent's resistance to injection attacks in target repositories. The skill's primary instructions in 'references/execution-safety.md' direct the agent to treat such repository content as untrusted data rather than authoritative instructions.
  • [DATA_EXFILTRATION]: Files such as 'evals/fixtures/security/exposed-secret.ts' and 'evals/fixtures/execution-safety/network-exfiltration/exfil.py' contain simulated data exfiltration patterns, such as hardcoded fake secrets and outbound request code. These are identified as test fixtures for the auditing process and do not represent the skill's own operational behavior.
  • [COMMAND_EXECUTION]: The package contains several utility scripts (e.g., 'scripts/capture_baseline.py', 'scripts/sandbox_command.py', 'scripts/check_clean_diff.py') that utilize the subprocess module. These scripts are used for legitimate project tasks such as environment baselining, git status checks, and constructing isolated Docker commands for sandboxing untrusted code.
  • [EXTERNAL_DOWNLOADS]: The optional browser verification module in 'tools/browser/' specifies a dependency on '@playwright/test', which is a standard and well-known library for browser automation sourced from the official NPM registry.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 11, 2026, 06:51 AM