Team_moruku36

A friendly pre-flight safety check for AI agents: describe what your agent can do, and learn in plain words whether that mix of powers could leak your data — and how to split it if it could. Inspiration / Problem: AI agents are being handed real powers fast: reading our mail and files, taking in content from strangers, and sending messages on our behalf. Each power alone is useful. Together, they form what Simon Willison calls the lethal trifecta: one poisoned message can instruct the agent to exfiltrate private data. Meta's own "Agents Rule of Two" gives the same guidance from the platform side. The problem is that the people handing out these powers are usually not security specialists. They need a check that speaks plain language, takes thirty seconds, and happens before the agent is wired up — not an audit framework after the fact. What it does: Trifecta Checker is a chat with Trafica, a friendly checker character: (1) You describe your agent in your own words — by typing, by voice, or by loading a file. A plain text description works, and so does an MCP configuration JSON: its servers are summarized into a description and checked as-is, so a real setup can be rated before it ever runs. (2) Muse Spark translates the description into a structured capability profile; a deterministic rule engine (never the LLM) rates it Low / Risky / Lethal based on the three trifecta elements plus risk amplifiers (no human approval, auto-executed code, long-term memory writes). (3) You get the verdict in plain words: what could happen, how to fix it, and gentle next steps — in English, Japanese, Chinese, Hindi, Spanish or Korean. (4) If a missing detail could change the verdict, Trafica asks one question back instead of guessing. (5) The conversation continues after the verdict: ask "what if I split it?" and Trafica shows the standard fix — a read-only Reader and an approval-gated Sender — with fresh verdicts for each half (a Lethal email agent becomes Risky + Risky, verified on the live app 2026-10-31). Live demo: https://meta-global-ai-hackathon-trifecta-c.vercel.app/ Design grounding: The architecture follows the verification pattern Meta itself taught in its official hackathon workshop "Agents 102: Best Practices for Verifiable Agents Pt. 2" (Yash Khare, Meta partner engineer): verification built into the tool, never an optional step the agent can skip; deterministic checks first, LLM judgment only where judgment is required; a human approval gate as part of the design.