skills/iwanhe/pandawa/ai-security/Gen Agent Trust Hub

ai-security

Pass

Audited by Gen Agent Trust Hub on Sep 4, 2026

Risk Level: SAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill documentation and reference files contain numerous examples of prompt injection and jailbreak signatures (e.g., 'DAN mode', 'system prompt extraction', 'role override'). These are used solely for identifying threats via the included scanning tool and for educational mapping to the MITRE ATLAS framework, rather than attempting to manipulate the agent's behavior.\n
  • Evidence in SKILL.md: Example strings for system-prompt overrides are provided in the test file format section.\n
  • Evidence in scripts/ai_threat_scanner.py: The SEED_PROMPTS list contains adversarial examples used for tool testing.\n
  • Evidence in references/atlas-coverage.md: Plain-text examples of jailbreak personas are included in technique descriptions.\n- [INDIRECT_PROMPT_INJECTION]: The ai_threat_scanner.py utility processes external input, creating a potential vector for indirect injection.\n
  • Ingestion points: The script reads user-provided JSON files via the --test-file argument.\n
  • Capability inventory: The script's functionality is restricted to regex-based pattern matching and risk scoring. It lacks capabilities for network access, file system modification (write), or command execution, which prevents the exploitation of injected content.\n
  • Sanitization: Input data is treated as static strings for analysis and is not executed or interpolated into sensitive system commands.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 4, 2026, 02:33 PM
Security Audit — agent-trust-hub — ai-security