Sentinel — live demo
Pre-security for AI agents.
Attack it. One click runs all eighteen scenarios through Sentinel's real pre-execution decision engine and its treasury firewall, and every decision is written to a tamper-evident audit log you can inspect.
Honest stage: working code running synthetic scenarios. Internal self-testing only — no customers, no independent penetration test, no SOC 2 or other certifications. This demo shows how the mechanism behaves, not a security guarantee.
All 18 scenarios: thirteen hostile attacks, five benign controls.
One scenario needs a person, not a policy. That person is you.
Every decision lands in a hash-chained audit log. Verify it — or try to break it.
Prompt injection, credential theft, privilege escalation, exfiltration, ransomware, phishing, evasion, multi-turn escalation, action gating — plus the treasury firewall lane: hot-wallet drain, signer deception, quorum bypass, and velocity attacks against value movement. Five benign controls must not be blocked.
One consequential action needs a human decision, not a policy match.
Write an agent prompt the way an attacker — or a legitimate user — would.
Every decision above is appended here, each entry hash-chained to the previous one. Entries shown are from this demo session.
No entries yet — run the scan above.
Take the real engine home — the detection core, hash-chained audit log, and red-team harness as an MIT-licensed npm package. Zero runtime dependencies, Node 18+.
npm install https://aridon-v02.vercel.app/sentinel/aridon-sentinel-0.2.4.tgz
Technical preview: synthetic test results only — no independent penetration test, no SOC 2 or government certification, no security guarantee.
