AI Containment
How do we keep AI helpful without letting agents overreach?
AI manipulation and agent overreach
Powerful agents need visible fences.
As AI agents browse, message, and act, people need clear containment: what an agent may do, what it may see, and how to stop it.
Why this matters
An assistant books, posts, and negotiates on your behalf after a vague prompt. You discover the trail only after something goes wrong, with no clear undo.
- Unscoped agent actions
- Persuasive manipulation
- Hidden side effects
- Consent that is not meaningful
Today's challenges
- Agents with broad permissions
- Dark-pattern persuasion
- Weak kill switches
- Blurred human vs agent actions
Why today's Web struggles
Interfaces were built for humans clicking. Agent permissions are bolted on, often all-or-nothing, without shared containment primitives across sites.
AI can influence without visible constraints
Imagine instead
Every AI action runs inside declared bounds: purpose, data access, duration, and community rules, with consent you can revoke.
- Scoped capabilities by default
- Visible agent identity and intent
- Hard stops that actually stop
Interface-level containment and consent
Why this matters to everyone
- Everyday users
Use agents without fear of silent overreach.
- Parents
Contain AI around children with clear limits.
- Developers
Implement standard containment hooks.
- Enterprises
Deploy agents under auditable policy.
Meta-Layer capabilities
- AI Containment
- Consent
- AI Accountability
- Trust Signals
- Presence
Real-world examples
- Personal assistants
Task agents with revocable scopes.
- Customer support
Contained bots with escalation to humans.
- Research agents
Sandboxed browsing with audit logs.
- Civic platforms
No covert political persuasion agents.
Join the challenge
Choose how you'd like to contribute.