magi
Pass
Audited by Gen Agent Trust Hub on Oct 22, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection because it ingests and deliberates upon external data, including user inputs and responses from external LLM engines.
- Ingestion points: User-provided decision framing in the
FRAMEphase and deliberation results from external subagents (Codex, Antigravity) in themultirecipe. - Capability inventory: Orchestration of multi-perspective deliberations and construction of prompts for authorized CLI-based analytical tools.
- Boundary markers: Implements a 'Perspective Independence Protocol' and 'Sequential Isolation Rule' to prevent cross-contamination and anchoring bias between viewpoints. An 'Ethics Gate' is enforced for expert channeling to prevent deceptive use.
- Sanitization: Employs structured JSON schema enforcement for subagent outputs and specific deliberation templates to constrain the reasoning environment.
- [COMMAND_EXECUTION]: The skill utilizes authorized command-line tools, specifically
codexandagy(Antigravity), to facilitate its 'Engine Mode' and 'Multi-Engine' deliberation features. - These tools are invoked using
codex execandagy -pwith dynamically constructed prompts derived from the local decision context. - The execution is scoped to analytical processing and is part of the skill's core consensus-building functionality.
- [EXTERNAL_DOWNLOADS]: The instructions reference strategic, financial, and industry benchmarks from well-known venture capital and technology organizations to ground its advisory modes.
- Documented sources include guidelines and datasets from Y Combinator, Sequoia Capital, a16z, Carta, and SFAI Labs.
- These references are used to provide evidence-based pattern diagnosis and financial modeling in 'Founder Mode' and 'Strategy Simulation'.
Audit Metadata