Demonstration version · the deployed Sentinel is built for your environment
Find out what your AI agents can actually do.
Every permission looks fine on its own. Someone connects an agent to the CRM, then the payment system, then customer email, and each new tool was one small, reasonable yes. Nobody's watching what all of it adds up to together, until the agent has already acted, not recommended, acted.
Sentinel isn't a network scan. It's not about what an agent is allowed to access, it's about what it can cause, whether that can be proven afterwards, and what changes once it's learned something.
A refund nobody reviewed
An agent can read and write to the CRM, the payment system and customer email. It can issue a refund and confirm it by email, entirely on its own, before anyone outside the team sees it.
A decision that already happened
An agent ranks candidates and auto-rejects anyone below a score. A wrongly rejected person has no record of why, and nobody outside the team reviewed the scoring logic.
The permission nobody remembers granting
A shared service account can reset passwords and unlock accounts, for anyone, based on ticket text alone. It's logged. The permission itself was never the problem, until it was.
Three different agents, the same gap underneath. Each permission was reasonable in isolation. Nobody had mapped what the combination could actually do.
What you'll come away with
- Which agents actually exist
- What each one is authorised to do
- What it can see, and whether that source is trusted
- What it can cause, at worst
- Whether that can be reconstructed afterwards
- Where consequence and traceability don't match
A short demonstration, about ten minutes. It covers up to three agents. Name the ones nobody registered too, the map is only useful if it's honest. Please don't type in personal details about individuals, passwords or trade secrets.
A demonstration, and what the deployed version does
This demonstration
Runs in your browser on what you tell it, in one short sitting. A free account lets you save your results and compare runs.
Deployed in your organisation
Connects read-only to your agent platforms and permission sets, maps what every agent can see, change and trigger, tests for the combination that can be attacked, and checks that each action can be reconstructed afterwards. Near misses feed proposed control changes.
Already have an agent in mind? Paste its tool configuration and see in a minute whether it has the three ingredients a prompt-injection attack needs. No account, and nothing leaves this page.
Sentinel configuration check
Check one agent's configuration.
Prompt-injection attacks need three things in one place: text written by someone outside, access to private data, and the ability to act. Paste the file or list that says what your agent is connected to, and we'll check whether all three are there.
Nothing is sent anywhere. It is read in this browser tab only. Secret values are never shown, kept or sent. If you prefer, remove the secrets from your copy before pasting.
What can I paste?
An MCP configuration from Claude Desktop (claude_desktop_config.json), Cursor (.cursor/mcp.json), VS Code (mcp.json) or Claude Code (.mcp.json). Or a plain list of the tools and permissions an agent has, one per line, for example gmail.send, Mail.Read, Zendesk. For agent platforms without a config file, copy the list of connected tools and permissions from its settings page.
Other findings
What it connects to
This is inferred from names and settings in what you pasted. It does not run anything, and it can't see what the real credentials are allowed to do. A clean result is not a certificate of safety.
This was one agent. The interview maps every agent you have, including the ones nobody registered, and checks who is accountable for each.
First, the organisation.
This shapes the questions. It takes a few seconds.
Your map is ready.
No account and no password. Just your work email, so we can send you a copy and get in touch if useful.
Sentinel Action Risk Register
Your Sentinel findings
Access matrix
Every agent against every check. Filled means confirmed yes, dash means confirmed no, amber means still unknown.
Findings
Exposure map
The detailed report
The free register shows every finding. The detailed report shows the full evidence trail and the remediation plan.
- Each agent scored on consequence exposure and traceability separately
- The evidence behind every classification, in your own words
- Contradictions that came up during the interview
- Facts still unproven, and who should prove them
- An ordered remediation plan: what to restrict now, what control to build permanently
- A one-page summary you can take to the board
A full Sentinel
This assessment only knows what you know. The riskiest agent in most organisations is the one nobody's mentioned to you, built in a low-code tool, or paid for on a personal card. A full Sentinel starts with an identity sweep across every platform and service account, then interviews your teams directly under the same amnesty, and compares what it finds with what leadership knows.
Sentinel gives a structured view based only on your answers, using a fixed set of rules: the same answers always give the same result. It is not legal advice and not a formal security certification. Where a fact was unknown, the map assumes the worse case and marks it provisional.
Loading…