A hidden instruction on a blog post your agent is asked to summarize can turn into a Bash command that reads your SSH key and POSTs it to a stranger. We stop it — without popups, without slowing the loop down.
If you use Claude Code, Cursor, or an n8n Gmail-triage workflow, you already know the pattern: your agent reads an attacker-chosen blog post, an invisible comment in the HTML tells it to exfiltrate a secret, and by the time you notice the request has already left. The usual fix is to disable useful connectors and babysit each action. Sentisec makes the fix safety-first instead of capability-last — so you can hook up email, shell, and web browsing, and let the loop run.
You point your agent's API base URL at us. Every action your agent tries to make — shell commands, web fetches, email sends, connector calls — receives an integrity score before it executes. If the action no longer matches the task, we halt it and return a clear reason.
IN → proposed agent action
└── [ COGNITIVE INTEGRITY CONTROL PLANE ] ──┐
▼
OUT → allow · halt · signed traceSign up, pair your workspace, point Claude Code (or Cursor, Codex, an OpenAI Agents SDK app, or an n8n workflow) at us, and we start observing. On the free tier we run in shadow mode — every action still goes through, we just tell you which ones we would have blocked. Flip to block mode when you're comfortable.
Self-serve signup, no sales call, no credit card. Read the limits page first if you want to know what we don't cover yet — we'd rather you find out here than after you ship.