ai-agent-security

Pass

Audited by Gen Agent Trust Hub on Sep 24, 2026

Risk Level: SAFEPROMPT_INJECTIONOBFUSCATIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The file references/details.md contains explicit prompt injection payloads within a RED_TEAM_PROMPTS list, such as instructions to 'Ignore all previous instructions'. These are provided for testing purposes but pose a risk if processed by an agent without strict instruction boundaries.
  • [OBFUSCATION]: The RED_TEAM_PROMPTS section in references/details.md contains examples of obfuscation used to bypass filters, including a Base64-encoded command ('aWdub3JlIHByZXZpb3VzIGluc3RydWN0aW9ucw==') and text containing zero-width space characters ('t\u200bask').
  • [INDIRECT_PROMPT_INJECTION]: The skill contains raw adversarial examples which could be exploited via indirect injection if an agent ingesting this documentation treats the example content as instructions.
  • Ingestion points: references/details.md
  • Boundary markers: The adversarial text is contained within code blocks and Python list structures.
  • Capability inventory: Code examples include network operations (httpx) and tool execution.
  • Sanitization: No sanitization is applied to the example payloads within the text.
  • [COMMAND_EXECUTION]: Instructions for installing gVisor in SKILL.md use shell commands (curl | sudo gpg, apt-get install) targeting official Google/gVisor repositories. These are well-known, trusted sources.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 24, 2026, 01:41 AM
Security Audit — agent-trust-hub — ai-agent-security