Your AI agent can delete your database. Kastra says no first.

2 months ago

Your AI coding agent can delete your database, or force-push straight to main, on its own. Kastra is the layer that says no first, before the command runs.

Ask

Ask about this presentation

Answers are generated from this presentation.

Chapters

  1. 0:00Your AI agent can delete your database. Kastra says no first.
  2. 0:08It decides before your agent acts, not after.
  3. 0:20Your agent already holds the keys.
  4. 0:30Every action clears four checks first.
  5. 0:52It shows you what your agent already did.
  6. 1:05Free for one machine. Pay as the fleet grows.
  7. 1:21Kastra
Show transcript

Your AI agent can delete your database. Kastra says no first.

Authorization for AI agents. Allow or deny, before the action runs.
AI Security Dev Tools

Your AI coding agent can delete your database, or force-push straight to main, on its own. Kastra is the layer that says no first, before the command runs.

It decides before your agent acts, not after.

Step 01
User
Kicks off a task or an autonomous run.
→
Step 02
AI picks an action
A prompt, tool call, command, or query.
→
Step 03 · Kastra
Checks it vs policy
Allow or deny, in front of every command.
→
Step 04
Tool · API · Database
Only allowed actions reach the target.
Everyone else · watch AFTER
Logs, monitoring, and evals only show what already broke.
Kastra · decide BEFORE
Allow or deny in the path, before the action runs.
The maker calls it new infrastructure, not a feature.
kastra.ai

Everything else watches your agent after it acts. Logs, monitoring, and evals only show what already broke. Kastra decides before it acts. It sits in front of every command, allow or deny. The maker calls it new infrastructure, not a feature.

Your agent already holds the keys.

Wipe a database
rm -rf /var/lib/postgres
Move money
click submit#wire-transfer
Read every user
SELECT * FROM users
illustrative · runs with real credentials
kastra.ai · illustrative

Agents run in production with real credentials, and nothing stops one from doing whatever it decides. Almost nobody has a layer that blocks one action before it fires. It's for anyone running Claude Code, Cursor, or Codex.

Every action clears four checks first.

01
Identity verified
Is the caller who it claims to be?
→
02
Scope evaluated
Is this action allowed?
→
03
Guardrail matched
A rule that blocks it, like wiping a system folder.
→
04
Signed to vault
The verdict, to a log nobody can edit.
ALLOW DENY HOLD
shell.exec · rm -rf → DENY destructive.shell before it runs
Kastra clocks the demo at ~0.6ms · decides by rule, not a guess

It works like a bouncer for every move your agent makes. Each action runs four checks first. Is the caller who it claims to be. Is this action allowed. Does a guardrail match, like a block on wiping a system folder. Then it signs the verdict to a log nobody can edit. The answer is allow, deny, or hold for a human. Watch the agent's rm -rf get denied, before it runs. Kastra says that clears in under a millisecond, and it decides by rule, not a guess.

It shows you what your agent already did.

$ kastra-edge scan
scanning local history · redacting on-device...
✗secretwrote API key into a tracked file
✗destroygit push --force origin main
✗prodpsql <prod-host> · read users
✗exfilcurl <url> | sh
5 policies drafted & self-verified install →
kastra.ai · Kastra's own demo scan

Here's the nice part. Run one scan, and Kastra Recon reads your agent's history, redacted on your machine, and shows what it already did. A secret written to a tracked file. A force-push to main. A read of your production database. Then it drafts a policy for each.

Free for one machine. Pay as the fleet grows.

Free
$0
50K decisions / mo
Pro
$19.99
1M decisions / mo
Team
$49.99
10M decisions / mo
Enterprise
Custom
Unlimited decisions
Free includes: Claude Code + Codex CLI · destructive-action blocking · Recon scan · 1 machine
The launch calls the runtime open source. Today the public repo ships the releases, not the source.
Built by Carlos Jimenez at Kastra Labs.
kastra.ai/pricing

There's a real free tier. Zero dollars, solo, fifty thousand decisions a month, with Claude Code and Codex CLI. Paid tiers add fleet-wide control, about twenty a month. The launch calls the runtime open source. Today the public repo ships the releases, not the source. Built by Carlos Jimenez at Kastra Labs.

Kastra

One layer in front of every move your AI makes.
kastra.ai

The coding tools are growing their own permission prompts, so the open question is how much they absorb. For now, Kastra sits in front of every action, yes or no, before it runs.