Vulnify (@vulnify) header
Vulnify (@vulnify) profile picture

Your AI agent can delete, export and send on its own. Vulnify is the approval step it's missing: allow, review or block. Free for devs.

Brazil
Joined May 2022
Photo by Vulnify (@vulnify)Photo by Vulnify (@vulnify)Photo by Vulnify (@vulnify)Photo by Vulnify (@vulnify)Photo by Vulnify (@vulnify)Photo by Vulnify (@vulnify)Photo by Vulnify (@vulnify)Photo by Vulnify (@vulnify)Photo by Vulnify (@vulnify)Photo by Vulnify (@vulnify)
Claude Code just added a setting that drops the last prompt bypass mode kept: the one before risky deletes. The problem was never prompts. It was asking about everything. Let routine calls run, block the never-ever ones, hold the risky middle for a person, in Vulnify or Slack.
1
1
2
208
What happens when nobody answers a review? It expires, and the final decision is BLOCK. Silence never turns into a yes. Your code runs the action only when finalDecision is ALLOW.
1
30
LangGraph: put one check in front of the tool node. The graph doesn't change shape. Reads run. A delete is blocked. An export to an outside address waits for an approver, then runs or not. The tool only runs when final_decision is ALLOW.
1
1
18
If people approve everything, approval is a policy bug, not a people problem. Routine calls should just run. The never-ever ones should be blocked without asking. Only the risky middle should reach a person.
1
10
Building your own guardrails for an agent is a fine place to start. Then you are maintaining the rules, the approval flow, the record of who approved what, and what happens when the check itself is down. We wrote down where DIY holds up and where it gets heavy.
1
15
Start your agent policies as a file, not a form. npx -p @vulnify/sdk vulnify init vulnify policies apply --dry-run vulnify test The CLI is free. Rules live in YAML and change in a reviewed PR. docs.vulnify.io/cli/?utm_sou…
42
decision is what Vulnify recorded. finalDecision is what your code obeys. Review pending: REVIEW Approved: ALLOW Denied or expired: BLOCK Run the action only when finalDecision is ALLOW.
1
62
3. How much at once? One record and 800 records should not get the same answer.
1
1
4. Who says yes when it's unclear? A person, with the reasons in front of them.
2
An agent with delete rights will eventually delete something it shouldn't. Not because it's malicious. Because "clean up the old rows" and "drop the table" look close enough to a model in a hurry. Make deletes a decision, not a permission.
2
1
79
If the docs and the API disagree, that's a bug. Here's the tour. Quickstart. Decisions and reviews. The risk engine. POST /v1/events and the gateway. Webhooks, SDKs, the CLI. Plans, and how your data is handled. API tables are generated from the public OpenAPI.
1
1
60