Before you hand an AI agent your inbox, your calendar, and your credit card — sign the paperwork. Marry the capability, protect the assets. Answer a few questions below and this tool drafts a plain-language permissions policy: what your agent may touch, what it may never do, who pays when it breaks something, and exactly how to shut it down. Til revocation do us part.
Key figure: 63% of organizations in IBM's 2025 breach study had no AI governance policy, and 97% of AI-breach victims had no proper AI access controls (IBM, Cost of a Data Breach Report 2025).
A prenup on paper is a promise. Agent Auth Gateway makes it enforceable — allowlisted tools, spend caps, approval queues, and a real kill switch, all enforced at runtime, outside the agent's reach.
Our platforms are being upgraded and return November 1 — get early access now, or explore the full Ingenious AI security suite:
Get Agent Auth Gateway early access →Only what its task genuinely needs — this is the least-privilege principle. Define the exact tools, data sources, and actions (read vs. write vs. send vs. delete) per task, start at read-only, and add broader access one deliberate step at a time. Never hand an agent blanket admin scopes "just in case."
Yes. Give every agent a distinct service identity instead of letting it inherit a human's credentials, and name one human owner accountable for what it does. If the policy lives where the agent can rewrite it, it is a suggestion — enforcement must sit outside the agent.
Run the kill switch in this order: revoke its API keys and tokens, disable its tool grants at the authorization layer, export the session logs for forensics, then verify no new tool calls occur after revocation. Use short-lived credentials so access expires on its own even if you forget.
The named human owner. Governance guidance is consistent on this: every agent needs an accountable individual who approved its permissions and scope before deployment. Document that ownership in the prenup — ambiguity is what makes incident response slow.
Keep stress-testing your agent setup:
Permissions, spending caps, data boundaries, and kill conditions. Define exactly which tools the agent may call, the maximum it may spend or change, what data it may read and share, and the signal that stops it. Write it before the agent touches production — renegotiating with a running agent means auditing everything it already did. The prenup turns ‘trust the agent’ into enforced limits.
A plain-language permissions policy you write before deploying an AI agent: what it may access, what it may never do, spend limits, who is liable when it breaks something, and the exact steps to revoke its access.
Only what its task genuinely needs. Define the exact tools, data sources, and actions per task, start at read-only, and widen access one deliberate step at a time — never blanket admin scopes.
Run the kill switch in order: revoke its API keys and tokens, disable its tool grants at the authorization layer, export the session logs for forensics, then verify no new tool calls occur.
The named human owner. Every agent needs one accountable individual who approved its permissions and scope before deployment — ambiguity is what makes incident response slow.