CANDIDATEREF A-4471
method GET
path /api/orders?id=1'||sleep(5)--
oracle awaiting replay
cvss unscored
01/ 05The offer
A finding
isn’t real
until it
crosses.
An AI pentest agent, run for you. It drives a real browser through your app like a user, captures every backend call, and lets nothing out of the red channel until a deterministic replay re-confirms it.
Book a callcal.com/himanshuat
Design-partner pilot · regulated and code-sensitive teams first.
Specificationrev. A
- Contained
- our own models, our own machines
- No third party
- no cloud AI provider sees your traffic
- Proof-gated
- replay-verified, or it never ships
- 44 checks
- injection · access control · exposure
- Output
- JSON · SARIF 2.1.0 · PDF · CI exit code
Authenticated testing. Scope allowlist, safe-mode, request budget and per-host rate limit on every run.
Scroll to cross the threshold
02/ 05The run
Everything that happens before the crossing.
- 01
Drive
A real Chromium session works the app the way a person does — form login, Google, or a saved session — clicking, submitting, paginating. Not a crawler hitting raw endpoints.
- 02
Capture
Every backend call it provokes is caught: XHR, GraphQL, WebSocket frames, form posts. The stack is fingerprinted passively from the same traffic and mapped to OSV and NVD advisories.
- 03
Probe
The captured surface is exercised by a differential engine and 44 checks. Blind bugs are proven out-of-band through an HTTP or DNS callback. Scope allowlist, safe-mode, request budget and per-host rate limit bound all of it.
Captured surfacesession A-44
- GET/api/session200
- POST/api/auth/login302
- GET/api/orders?id=41200
- POST/graphql {orders(first:50)}200
- GET/api/users/me200
- GET/assets/app.<hash>.js.map200? candidate
- WS/socket order.updated101
- POST/api/orders201
- GET/api/orders?id=41%27500? candidate
- GET/api/orders?id=1%27||sleep(5)--200→ REF A-4471
candidates raised 3confirmed pending verification
Go control plane · Playwright browser sidecar · open-weight models we host ourselves.
Nothing is sent to a third-party model provider. Every prompt, tool call and request logged.
03/ 05The crossing
Red channel
Something in the response looked wrong. No proof. Never counted, never scored.
Green channel
Re-confirmed by replay. Scored, tagged, proof attached. The only thing that reaches your report.
replayoraclescore
04/ 05What it is
01
Continuous
Runs on every deploy instead of once a year. Exposure shrinks from months to hours.
02
Proven
A candidate→confirmed gate, not a probability score.44 replay-confirmed checks · ~90 orchestrated tools
03
Contained
The agent runs on our own infrastructure, on open-weight models we host ourselves. No cloud AI provider ever receives your traffic.Run data retained for re-runs · deleted on request