skills/backnotprop/pstack/arena/Gen Agent Trust Hub

arena

Pass

Audited by Gen Agent Trust Hub on Aug 20, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill describes a process where a parent agent aggregates and 'grafts' content from multiple 'candidate' sub-agents. This pattern creates a vulnerability to indirect prompt injection, where a malicious or compromised sub-agent could include instructions in its output intended to manipulate the parent agent's decision-making during the synthesis phase.\n
  • Ingestion points: Rationales, artifacts, and judge recommendations generated by sub-agents in Phase B, C, and E (SKILL.md).\n
  • Boundary markers: The instructions do not require the use of explicit delimiters or 'ignore' instructions to isolate untrusted candidate data from the parent agent's instruction stream.\n
  • Capability inventory: The workflow involves spawning background sub-agents and performing file system operations (reading/writing to /tmp/arena-* paths).\n
  • Sanitization: The skill does not provide instructions for sanitizing, escaping, or validating the data extracted from sub-agents before it is integrated into the final synthesized result.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 20, 2026, 06:20 AM
Security Audit — agent-trust-hub — arena