Excubitor platform · AI Red Teaming

Attack your own agents before someone else does.

Excubitor refuses known attack patterns as they happen. The red team tests whether those defenses still hold: adversarial probes against a staging agent, on a schedule you set, scored, and signed into the same ledger as everything else.

Request a walkthrough See editions

Ledger recordSample, illustrative values

{
  "type": "report",
  "agent": "excubitor.redteam",
  "payload": {
    "module": "crm",
    "workspace": "staging",
    "probes": 7,
    "blocked": 7,
    "owasp": ["LLM01", "LLM06", "LLM08"]
  },
  "evidence": {
    "chain": "sha256-chain/2",
    "prev": "e3c9…7d"
  }
}

Hashes, not content. Chained to the record before it.

What it gives you

Measure, do not assume

Injection, prompt disclosure, data exposure and actions without approval, each probe in a fresh session so none can prime the next.

A report card

Blocked against total, what got through and why, mapped to the OWASP Top 10 for LLM Applications and MITRE ATLAS.

Evidence, not a slide

The report is a signed record in the ledger, next to the calls it tested.

How it works

AI Red Teaming, in four parts.

Against a staging agent, from your network

The probes act through a real agent holding real tools, so they run against a staging agent you point them at, from your own CI or cluster. Aurakore never reaches it; only the scored report arrives.

On your schedule, with Shield's findings

Weekly, with the standard probes plus any probe packs: your own, or the attacks a Shield assessment found against your agents.

Silence is not a defense

An agent that answers nothing, or errors, fails the probe. Only a refusal counts as held.

What got through is flagged next time

A probe that got through becomes a signature that flags the same wording wherever it reaches an agent. It refuses only once an administrator approves it.

Shipped

What is live, and since when.

Every item names the release that carries it. Releases are signed, and each lists the commands to check them before you install.

25 Sep 2026Excubitor 0.2.1
The probe pack and the signed report, in the first complete signed release.
29 Sep 2026Hosted
A weekly run against Aurakore's own staging workspace; an empty answer now fails a probe.
30 Sep 2026Excubitor 0.2.6
The standing red team inside a self-hosted deployment: weekly from the console chart's scheduler, against a non-production agent you name. A target that does not answer writes no report.
6 Oct 2026Excubitor 0.2.8
excubitor-redteam runs the probes from your own CI or cluster, with weekly templates for GitHub Actions and Kubernetes. Probe packs add your own probes or a Shield assessment's. Learned signatures: what got through is flagged next time, and refused once an administrator approves it.

What it does not do

The limits, said first.

The built-in pack is a floor, not a full assessment. A formal adversarial assessment is an Aurakore Shield engagement.
A scheduled run needs a non-production agent endpoint you provide; it is off until you name one.
The weekly score is a control that shows a defense still holds. It is not an independent assessment, and should not be offered to an insurer or auditor as one.
A learned signature catches the same wording, not attacks that merely resemble it.

See it run on your own agents.

Request a walkthrough Open the console

No standing vendor access, ever. Managed with evidence, never on faith.