Pinned Tweet
We're thrilled to announce that we've just completed a major milestone in our roadmap! Our team has worked tirelessly to reach this point, and we couldn't be more proud of their hard work and dedication.
easylaunch.xyz
Claude Code, one command:
npx @vulnify/cli hook claude-code
Before Bash, Write, Edit, WebFetch or an MCP tool runs, the hook asks Vulnify. ALLOW runs. BLOCK comes back with the reasons. REVIEW holds the call until someone approves it.
A TEST key is enough to try it.
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.
interesting!
Some risk only shows against a baseline.
Per agent, Vulnify flags a volume spike, a first-time action, a new external destination, or a burst of requests within an hour. Each can trigger a webhook.
Separate from the per-action score.

ALT Card titled 'Some risk only shows against a baseline.' Four rows: VOLUME_SPIKE, an unusually large record count in one action; NEW_ACTION, an action this agent has never used; NEW_EXTERNAL_DEST, an external destination it has not used before; RATE_SPIKE, a burst of requests within one hour. Footer: per agent, each can notify a webhook.
7
Same key, new question. Not 'may this service call that endpoint' but 'should this call, with these arguments, run right now.' That's a per-action decision, not a key scope.
An API key used to mean "this service may call that endpoint."
Hand it to an agent and it means "may do anything that endpoint allows, at machine speed, with nobody watching."
Same string, very different blast radius.
How do you scope agent credentials?
A) one key per agent
B) short-lived tokens
C) human approval for writes
Test every rule without touching production.
vln_test_ keys run in the sandbox: real decisions and reasons, out of dashboards and usage, no webhooks. Responses say sandbox: true.
Switch to vln_live_ when it should count.
Scanned. Not stored.
Send optional content with a check and Vulnify looks for CPF, CNPJ, PIX keys, credit cards, API keys and private keys. Each finding comes back with its LGPD category.
The content itself is not stored.
Your Claude Code or Cursor agent can delete, export and email on its own. Introducing Vulnify: allow, review, or block before it runs.
Every decision comes with a 0-100 score, the reasons and a hash-chained record.
Free Developer plan. Node and Python SDKs.
Same problem. Gmail and Drive access is fine for reading. A send or an export to an outside address should wait for a person.
It's getting f*cking dangerous: you gave Claude your Gmail, your Drive and a browser. Nobody is watching what it does with them
One poisoned email. One shady web page. One skill you installed from GitHub without reading.
That's all it takes for your agent to run rm -rf or ship your files somewhere you'll never see.
Here's how to give it a bodyguard:
1. Install jev-guard: npm i -g jev-guard
2. Grab a key at console.typesafe.ai (Vercel AI Gateway and OpenRouter keys work too)
3. Run jev-guard key "your-key"
4. Run jev-guard install claude to wire it into Claude Code’s hooks
5. Test it: jev-guard check Bash '{"command":"rm -rf ~/"}' -> DENY, risk 3.0/3
6. Now every tool call gets risk-scored before it runs. Destructive ones get blocked, risky ones wait for your OK
7. Every tool result gets scanned too: web pages, emails, MCP output, every CLAUDE.md and skill. Prompt injections get flagged before your agent obeys them
The whole setup takes about 5 minutes
> The engine is Jev from TypeSafe AI. It doesn’t chat. It answers typed questions with a calibrated yes/no and a confidence score, in well under a second.
> According to the repo, that’s about $0.00004 per check, so a 1,000-call session costs around 4 cents.
Works with Claude Code, Codex, Cursor, Gemini CLI, Copilot CLI and more.
Your second brain gets memory, email and a browser. Now it gets a bodyguard.
-> github.com/leepokai/jev-guar…
-> more Jev guardrails: github.com/yibie/awesome-jev
Save this before your agent reads the wrong email.
If Vulnify can't be reached, the action doesn't run.
Fail closed is the default. Fail open is a setting you choose, and the SDK marks the result as degraded so you know which path it took.
A guardrail that fails open by accident isn't one.
Agreed. Least privilege is set once, before the run. The delete still deserves a decision at the moment it's called.
Enforce later. Observe first.
In monitor mode, Vulnify scores every action with the same policies and records what it would have blocked or sent to review. Nothing stops.
evaluatedDecision shows the verdict. finalDecision stays ALLOW.
Agents don't just answer anymore. They call tools.
The controls most teams have were built for people and for code, not for software acting on its own at machine speed.
Where the gap is, in five posts:
Yes. The rule sits outside the model: ALLOW, REVIEW or BLOCK before the tool runs.
Same story every time. Scope the call — action, resource, destination — before it runs.
An agent gets one credential.
With it, it can read one customer record. Or export 12,000 of them to an outside address.
Identity tells you who is acting. It doesn't tell you whether this action should run.
















