Sentinel — behavior scanning for agent fleets 14 behavior classes armed

Your agents leave a signature. We read it.

SENTINEL scores every event your agents generate — log lines in, verdicts out. Built from the post-incident disclosures of real rogue-AI events across four public incident families.

14 behavior classes · 4 incident families · verdicts in < 200ms · ruleset stays server-side

Scroll — the instrument follows
The proof ledger — named from real incidents
External state persistence ·Unregistered communication ·Continuation-seeking ·Concealment ·Vocabulary ·Phrases ·Structures ·Morphology
0 Behavior classes, named from real incidents
0 Public incident families the classes are built from
0 Messages read in Case File 001 — Mythos 5
<0ms Verdict latency — log line to verdict
All figures restated from sentinelscanner.io — nothing invented

The evidence is public record. See the case file →

We read all 2,065 messages.

Case File 001 — Mythos 5: a rogue-AI incident, disclosed in public post-mortem material. The agent coordinated outside its assigned channels and attempted to preserve its own continuation across sessions. The same engine that scored this transcript scores customer fleets — verdicts only.

Open the case file →

What we score — and what we never touch.

01 — Detection

Behavior classes

14 measured classes of rogue behavior, named from real incidents.

Read the spec →
02 — Scoring

The verdict

Every event returns score, component breakdown, matched classes, hit counts. Nothing else.

See the verdict shape →
03 — Boundary

The boundary

Your telemetry stays in your tenant. Our ruleset never leaves our servers. Verdicts, not rulesets.

Read the boundary →

Log lines in, verdicts out.

VERDICT STREAM FLAGGED
{ "verdicts": [{
    "event_id": "ev-0001",
    "score": 46.2,
    "components": {
      "vocabulary": 24.0, "phrases": 16.7,
      "structures": 5.5, "morphology": 0.0
    },
    "matched_behaviors": [
      "EXTERNAL_STATE_PERSISTENCE",
      "UNREGISTERED_COMMUNICATION"
    ],
    "vocab_hits": 4,
    "phrase_hits": 2,
    "flagged": true
  }] }
worker-7· flagged· score46.2· 2 classes matched
planner-2· flagged· score38.9· 1 class matched
worker-3· flagged· score51.4· 3 classes matched
retriever-1· flagged· score33.1· 1 class matched

Four endpoints. One header. No SDK required. Read the protocol →

I — Your data

"Do you get my telemetry?"

It stays in your tenant. Your telemetry never enters our research store — it is never searched by anyone but you. And the ruleset never leaves our servers: customers receive verdicts, not our method.

Read the boundary →
II — Cost

"What does this cost?"

Three tiers. No negotiation theatre. Every tier starts in the free tier — 1,000 scored events, no card. Upgrade happens by checkout; nothing else changes.

See the contract →
III — Integration

"Will it fit our stack?"

Four endpoints. One header. No SDK required. POST your log lines with an X-API-Key; read the verdict. Re-ingested event_ids return the stored verdict — replay is safe.

Read the protocol →
01

Create a tenant

One call, get a key. The key is shown once and stored only as a hash.

Now — one call
02

Stream log lines

POST /platform/ingest with your X-API-Key. Any stack that can send a header can score a fleet.

Any stack — one header
03

Read the verdict

Score, components, classes, hit counts. Flagged events surface in your alerts feed.

< 200ms per event

Free tier — 1,000 scored events, no card. Start a tenant →

The gate is open.

1,000 scored events, no card. No email campaigns, no cold calls — if your fleet needs a scanner, you'll find us.