self-improve

Warn

Audited by Gen Agent Trust Hub on Nov 3, 2026

Risk Level: MEDIUMINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests and processes all files within a target repository to generate research briefs and plans. This creates a surface for indirect prompt injection where malicious content in the repository (e.g., hidden directions in comments or documentation) could influence the agent's logic or execution.
  • Ingestion points: The researcher agent (si-researcher.md) reads the entire directory specified by repo_path.
  • Boundary markers: There are no explicit markers or instructions provided to the agent to treat codebase content as data rather than instructions.
  • Capability inventory: The system can write to the filesystem (Git), execute shell commands (benchmark_command), and perform network operations (Git push).
  • Sanitization: No sanitization is performed on the repository content before it is passed into the agent's context.
  • [DYNAMIC_EXECUTION]: The si-benchmark-builder agent is responsible for creating a benchmark script (likely in Python) within the target repository. This script is then executed by the executor agent. This pattern of generating and then executing local code based on autonomous analysis is a significant dynamic execution risk.
  • [COMMAND_EXECUTION]: The orchestrator autonomously performs Git operations (worktree management, merges) and executes user-confirmed benchmark commands. This execution continues over many iterations without re-confirming with the user.
  • [DATA_EXFILTRATION]: The skill contains logic to perform git push operations to the repository's configured origin. If the agent's planning phase is compromised via indirect injection, it could be manipulated into committing sensitive files from the local environment and pushing them to a remote server.
  • [PROMPT_INJECTION]: The orchestrator instructions explicitly command the agent to 'NEVER stop or pause to ask the user' and to 'Trust' benchmark commands. This essentially overrides the standard human-in-the-loop safety protocols for the duration of the execution.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 3, 2026, 06:06 AM