- Section
- AI security
- Ref
- BM-AI
- Updated
The AI answers to you.
Two things on one sheet: how BreachModal’s own AI is kept on a short leash, and how we test and harden the AI features you are shipping.
§1How our AI is held to account
It proposes. A person signs. The record keeps both.
The same six entries appear in every incident log, in this order, with no step skipped.
- T+00:00:04SYSTEM
Observes
The AI reads telemetry with read-only access. Looking changes nothing.
- T+00:02:51AI
Proposes
A plan arrives with its evidence, its confidence, the reason in plain language and the steps to roll it back.
- T+00:03:12ANALYST
Reviews
A named person reads the proposal and the evidence behind it. They can change it or reject it.
- T+00:03:40ANALYST
Signs
Approval takes a full, deliberate signature. A half-press does nothing. Nothing happens without it.
- T+00:03:41SYSTEM
Executes in scope
Only the signed action runs, with tool grants limited to what that action needs.
- T+00:04:41SYSTEM
Seals
Every step, the evidence and the signature are logged and hashed into the evidence packet.
§2Threat model for your AI
Four findings we look for first.
Ship AI features with guardrails: inputs and outputs checked against policy, data paths kept apart, abuse resistance you can measure.
| Ref | Finding | How we test it |
|---|---|---|
| F-1 | Prompt injection and data exfiltration | We plant instructions in documents, tickets, web pages and tool output, and try to make the model leak data or take actions it should not. |
| F-2 | Model and embedding supply chain | We trace where weights, embeddings and plug-ins come from, and check what happens when one of them is tampered with. |
| F-3 | Tenant isolation and vector-store leakage | We try to retrieve another tenant’s documents through search, retrieval and shared caches. |
| F-4 | Abuse, jailbreaks and missing monitoring | We run jailbreak suites and misuse scenarios, then check whether anyone would have noticed. |
§3 · Procedure
We attack it.
Systematic jailbreaks, tool-misuse scenarios and automated fuzzing across prompts and tools, run against your real configuration rather than a demo.
Then we fence it.
Policy filters, isolation boundaries, least-privilege tool grants and monitoring after deployment, each one re-tested before we sign it off.
Before you ship
Have your AI tested before someone else does.
Tell us what your model can read and what it can do. We will scope a red-team exercise around exactly that.