Hi Product Hunt 👋 I'm Giovanni, and I build Vulnify.

If your Claude Code, Cursor or Grok agent can delete, export or send email, Vulnify decides each of those actions before it runs. Free plan, no card.

The problem: agents get a credential once and then act on their own. The credential says who is acting. It doesn't say whether this export, of this many records, to this destination, should run right now.

How it works: your code describes the action, and Vulnify returns ALLOW, REVIEW or BLOCK, with a 0-100 score and the reasons. REVIEW waits for a person, in the app or in Slack. BLOCK never starts. Every decision lands in a hash-chained audit log.

A few things I chose on purpose:

Fail closed by default: if Vulnify can't be reached, the action doesn't run unless you opt into fail open.
Monitor mode: see what would be blocked before anything is.
Metadata, not data: Vulnify sees the agent, action, destination and record count, never the records.
Policies as code: a CLI validates, dry-runs and tests YAML policies in CI.

It's early access. Not there yet: SSO, enforced org-wide MFA, a dedicated environment, certifications. The SDKs are MIT on GitHub.

One question for you: how do you decide today which agent actions need a human? And where would this model not fit your stack?

Security contact: security@vulnify.io (https://www.vulnify.io/disclosure)