mantis-meta-agent

Pass

Audited by Gen Agent Trust Hub on Sep 14, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • Automated File Management: The skill performs file system operations including directory creation, moving finding reports to archives, and deleting temporary backup files. The deletion logic uses a specific UUID-anchored regex to ensure only files created by the patching process are removed, avoiding user data or critical state directories.
  • Version Control System (VCS) Integration: The supervisor interacts with Git, Mercurial, and Repo tools to synchronize the target project with its upstream and detect the current state. These operations are gated by 'dirty tree' checks to prevent overwriting local uncommitted changes.
  • Snapshot and Immutability Management: To ensure analysis consistency, the skill creates immutable snapshots of the target code using tools like git worktree or directory copying. It applies read-only permissions (chmod -R a-w) to these snapshots to prevent accidental modification during the review loop.
  • Indirect Prompt Injection Surface: As an orchestrator, the skill ingests findings generated by subagents that audit external code or binaries. While this introduces a surface for indirect prompt injection—where malicious patterns in a target codebase could influence finding metadata—the skill employs status verification (e.g., checking patch_status and vcs_info) and isolates auditing tasks to specialized subagents to mitigate risk.
  • Subagent Delegation: The meta-agent orchestrates a multi-stage pipeline by delegating tasks to specialized subagents (e.g., @mantis-researcher, @mantis-patch). This modular approach keeps auditing context separate from the high-level supervision logic.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 14, 2026, 05:43 PM
Security Audit — agent-trust-hub — mantis-meta-agent