Safety, plainly

Everything we do to keep MagnusAI from doing the wrong thing on your behalf.

Green-light consent

Every permission is off until you flip it on. Four preset modes (Cautious, Balanced, Autopilot, Custom). In-the-moment prompts before any irreversible action — voice or in-app modal. Read the consent doc.

Audit log preview (V1-F5)

Every action MagnusAI takes is tappable. You see exactly why: which capabilities were considered, which scopes were checked, which user instructions led here.

Posted to TikTok @ 6:00pm

Why: you said "share my new drop everywhere at 6pm" on Apr 24. Reversibility: irreversible.

Tipped 3 commenters $5 each ($15 Cred)

Why: you said "tip the top 3 commenters from yesterday's stream". Under voice-auth threshold ($100).

Voice authentication (V1-F6)

Set a $-threshold (default $100). Any action above it requires a voiceprint match. Friend grabs your phone? They can't spend your money.

User-set rate limits (V1-F7)

Tell MagnusAI "don't let me spend more than $100 on ads this week even if I tell you to." Commitment device.

Anti-sycophancy (V1-F4)

"Push back when the user is about to make a costly mistake. Be direct, not servile. Useful, not agreeable. If the user's plan has a flaw, say so before proceeding."

Baked into MagnusAI's system prompt. Trusted advisor, not a yes-machine.

"What I can't see" disclosure (V1-F10)

When you ask MagnusAI about a capability it doesn't have, it names the gap explicitly: "I can see your calendar but not your email — want to connect Gmail?" Honesty over omniscience.

Minors policy (V1-F2)

Age verification at signup. Separate minor-mode with restricted capabilities. No voice cloning, no purchases, no social posting for minors without parental consent. Stricter under-13 per COPPA.

Lawyer review required before launch — see // LEGAL: lawyer review required before launch — minors handling markers in code.

Kill-my-data (V1-F1)

One click in /app. Audit logs exported, all connected accounts disconnected, account hard-purged on a 7-day window. GDPR / CCPA compliant.

Emergency stop

Single voice command ("MagnusAI: stop.") or button in the app pauses every in-flight action instantly. Resume requires explicit re-grant.

Capability gating (legal safety)

● GREEN — auto-allowed
● YELLOW — regulated; written sign-off + audit trail
● ORANGE — gray intent; logged, never built
● RED — illegal; auto-rejected
? UNKNOWN — escalates with research summary

Money, minors, medical, firearms, privacy, gambling, securities — always YELLOW or harder. Lawyer-consult always.