Prompt rules
"Never run destructive commands." The agent in that incident had a rule like this, broke it, and then admitted it.
Early access · open-source broker
Deadbolt holds them. Your agents keep reads and safe writes. Anything you can't undo, like dropping a database, deleting a volume or changing IAM, waits until an independent reviewer (and for the scary ones, a human) signs that exact action.
aws rds delete-db-instance
--db-instance-identifier prod-main
--delete-automated-backups
The nine-second problem
In a reported April 2026 incident, a coding agent fixing a staging issue found a broad API token and deleted a startup's production database and its backups in one call. There was no confirmation step and no second key. Here's a simulated version on AWS, with and without Deadbolt.
# press play
The replay is a simulation. With Deadbolt on, the receipt is signed with real Ed25519 keys generated in your browser. The replay makes no network calls.
Why guardrails fail
"Never run destructive commands." The agent in that incident had a rule like this, broke it, and then admitted it.
After enough "Allow?" prompts, people click without reading. And the agent still holds the credential if anything goes around the prompt.
Whoever holds it can do anything. Often, that's every agent you run.
How it works
Agents get a scoped Deadbolt token. The privileged AWS role and the Postgres admin user live in the broker, in your VPC.
Opinionated defaults on day one: reads pass, deletes and IAM changes are critical. Edit it all in one YAML file.
Review tier: an agent from a different model vendor signs the exact action digest. Human tier: on-call approves in Slack with a passkey. Critical: both.
The broker runs the action with its own role and writes a signed receipt: who asked, who checked, what exactly ran.
| pass | Describe, Get, SELECT | direct, no broker |
| review | Scoped updates, instance changes | cross-vendor agent signs |
| human | IAM, ALTER | passkey in Slack |
| critical | Delete DB / volume / bucket, DROP, TRUNCATE | both |
# deadbolt.yaml protect: aws: { role: arn:aws:iam::…:role/deadbolt-privileged } postgres: { dsn_env: PROD_DB_ADMIN } tiers: pass: ["aws:Describe*", "sql:SELECT"] review: ["sql:UPDATE rows<1000"] human: ["aws:iam:*", "sql:ALTER"] critical: ["aws:rds:DeleteDBInstance", "aws:ec2:DeleteVolume", "aws:s3:DeleteBucket", "sql:DROP", "sql:TRUNCATE"] reviewer: must_differ_from: author_vendor key: env:REVIEWER_API_KEY # bring your own approvers: slack: "#prod-approvals" passkey: required
Your keys stay home
The broker will be open source and runs in your VPC. Our cloud sees approval requests, signatures and receipts. It never sees credentials or data.
Compared
| Deadbolt | Prompt rules | Harness prompts | Enterprise access platforms | |
|---|---|---|---|---|
| Agent holds the prod credential | Never | Yes | Yes | No |
| Independent, cross-vendor reviewer | Required | — | — | Varies |
| Signed, portable receipt | Every action | — | — | Audit log |
Early access is free while we build it. Leave your email and we'll be in touch.