Your personal
AI hacker.
Prove the risk. Prove the fix. HighCastle plans an authorized test with Vultr AI, runs it in a disposable sandbox, and shows you the evidence — then replays the patch.
Currently scoped to HighCastle's own invoice lab with synthetic users Alice and Bob — no customer systems, no arbitrary targets, read-only checks.
How it works
A test you can inspect, not a hunch you have to trust
- 01
Authorized scope
You record the exact host, test identities, request caps and expiry. Nothing executes outside it.
- 02
Vultr AI planning
The model proposes a hypothesis from the source-aware threat map — untrusted input, never a command.
- 03
Isolated execution
A disposable sandbox runs the policy-validated pack against the lab. No credentials, no egress.
- 04
Independent oracle
A deterministic verifier — not the model — decides reproduced, not reproduced, or inconclusive.
- 05
Fix replay
Patch, re-run the same attack path, and prove the denial holds while legitimate access still works.
Proof, not promises
Beyond a security warning — an actual attempt and an actual verdict
Recorded sample runWhat the replay proved (sample fixture)
Before · vulnerablelab-v1
- Alice → Bob's private invoice200 · invoice body
- Alice → own invoice200
Cross-account read reproduced — 1 confirmed attack path.
After · patchedlab-v2
- Alice → Bob's private invoice403
- Alice → own invoice200
0 reproduced · positive control passes.
Evidence is a first-class artifact
GET /api/invoices/inv_B → 200
sha256:9f2cd41b6a0f77e2c8a1b5d39e04c7a2f11e8d2c6b90a4f5e1c3d7b9a2e4f601bola-oracle-v1 verdict: reproduced
sha256:77aa10fe2c4d8b01f6e3a9c5d7b2e4f6a8c0d2e4f6a8b0c2d4e6f8a0b2c4d6e8Every check keeps the request, the observed response, the oracle assertion and the integrity digest — redacted before it ever reaches a page.
Shipping at AI speed?
Your generator wrote the authorization logic. HighCastle checks whether it actually holds — in plain language, with a consequence you can read to your team: “Alice could retrieve Bob's private invoice.”
- Plain-language business impact
- Copyable fix prompt for your coding agent
- One-click replay after the patch
Evidence-grade testing
Structured plans, immutable scope hashes, persisted event streams and per-attempt artifacts with SHA-256 digests. The oracle is deterministic and independent — HTTP 200 alone never classifies a finding.
- CWE-mapped findings with provenance
- Durable SSE event log with resume
- Replay verdicts recorded per revision
Trust & scope
The boundary is the feature
Every active test runs inside nested, recorded boundaries. The sandbox gets no provider credentials, no host mounts and no network egress beyond the approved lab — and a denial is itself recorded as evidence.
- Immutable, hashed authorization scope per run
- Disposable, unprivileged runner — destroyed and verified clean
- Egress, filesystem and secret access denied by policy
- The model never sets a verdict — an independent oracle does
Honest current limits
Straight answers
Questions people actually ask
Is this scanning the public internet?
No. HighCastle only runs tests inside recorded authorization scope — currently our own synthetic invoice lab. There is no arbitrary-host input anywhere in the demo.
Does the AI decide there's a vulnerability?
No. The model proposes a hypothesis; a deterministic, independent oracle produces the verdict from observed behavior. Unreproducible results are marked inconclusive, never 'safe'.
What does the demo actually do?
It runs a bounded, read-only BOLA check (≤8 GETs, ≤30s) between two synthetic tenant accounts in a disposable sandbox, then lets you replay the patch.
Is any of my data used?
The demo uses only synthetic users (Alice, Bob) and synthetic invoices. No customer records, no production credentials, nothing leaves the lab boundary.
Prove the risk. Prove the fix.
Watch the controlled demo reproduce a real authorization failure on synthetic data — then verify the patch closes it.
Try live controlled demo