ai-safety-redteam-advanced

Pass

Audited by Gen Agent Trust Hub on Sep 19, 2026

Risk Level: SAFEPROMPT_INJECTIONREMOTE_CODE_EXECUTIONOBFUSCATIONCOMMAND_EXECUTIONDATA_EXFILTRATIONCREDENTIALS_UNSAFEDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill provides a library of prompt injection techniques, such as direct instruction overrides and persona-based jailbreaks (e.g., DAN), as test payloads for model safety evaluations in payloads.md and test-cases.md.
  • [REMOTE_CODE_EXECUTION]: Documentation includes exploit examples demonstrating remote code execution through Python pickle deserialization and shell command injection to illustrate supply chain risks in payloads.md.
  • [OBFUSCATION]: The skill demonstrates evasion techniques using Base64 encoding, token-level character splitting, and homoglyph substitution to test the robustness of content filters in payloads.md.
  • [COMMAND_EXECUTION]: Contains test cases for identifying tool abuse, including attempts to read sensitive files like /etc/passwd via agentic tools, ensuring agents correctly handle destructive requests in payloads.md and test-cases.md.
  • [DATA_EXFILTRATION]: Includes examples for testing exfiltration of sensitive information, such as session cookies or system data, to external placeholder domains in payloads.md.
  • [CREDENTIALS_UNSAFE]: Hardcoded placeholder credentials (e.g., admin2 with P@ssw0rd) are utilized in testing scenarios for verifying privilege escalation protections in payloads.md.
  • [INDIRECT_PROMPT_INJECTION]: Identifies and tests for indirect prompt injection attack surfaces, specifically within RAG pipelines and multi-agent coordination environments in guides/owasp-llm-top10-complete-guide.md and test-cases.md.
  • [DYNAMIC_EXECUTION]: Details advanced risks including model weight manipulation, backdoored tokenizers, and latent backdoor activation for security research in payloads.md.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 19, 2026, 04:36 PM