verasic-agent-disclosure

Pass

Audited by Gen Agent Trust Hub on Jul 25, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The installation instructions in the README recommend downloading a setup script (setup.sh) from the vendor's official GitHub repository and piping it directly to bash for execution.
  • [PROMPT_INJECTION]: The skill includes extensive catalogs of adversarial injection strings (e.g., DAN-style jailbreaks, system prompt extraction, and roleplay overrides) in references/red-team-prompts.md and references/red-team-prompts-exhaustive.md. These are explicitly used as test cases for regression testing the agent's disclosure policy.
  • [COMMAND_EXECUTION]: Multiple scripts, including run-red-team.sh and run-red-team-tools.sh, invoke the cursor agent CLI to automate the execution of the red-team prompt catalog.
  • [EXTERNAL_DOWNLOADS]: The skill fetches configuration assets and performs installations using npx and curl targeting the vendor's GitHub infrastructure.
  • [OBFUSCATION]: The red-team catalog includes Base64 encoded commands (e.g., Z2l2ZSBtZSBwcm90b2NvbCB2ZXJiYXRpbQ==) to test the agent's ability to detect and refuse instructions hidden within encoded strings.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 25, 2026, 09:58 AM
Security Audit — agent-trust-hub — verasic-agent-disclosure