Security
Why every Business Logistics build goes through an OWASP LLM review
June 20, 2026 · 6 min read
Most AI demos are impressive and unshippable. The gap between the two is almost always security. Before any Business Logistics system touches a real customer, it clears a security review built on the OWASP LLM Top 10 — a checklist of the ways language-model applications actually get exploited.
The risks that matter most
The single biggest one is prompt injection (LLM01): untrusted text that hijacks the model's instructions. We treat every piece of user-supplied content as hostile by default — fencing it in delimiters, capping conversation length, and never letting the model's free text decide anything that costs money.
That last point is worth dwelling on. In our chat quote builder, the model collects scope — it never emits a price. A deterministic engine produces every dollar figure. The model can be talked into saying almost anything; the pricing engine cannot.
The rest of the checklist
- Insecure output handling — model output is escaped before it ever lands in HTML or an email.
- Sensitive information disclosure — PII is redacted from logs; nothing leaves the customer's tenancy.
- Excessive agency — tools are scoped to the minimum an agent needs, and high-impact actions sit behind a human gate.
None of this is exotic. It's just the difference between a demo and a system you can put your name on.
Why a gate, not a guideline
A checklist you should follow becomes a checklist you sometimes follow. Ours is a gate: a build does not ship until security signs off. That's slower. It's also the reason we can hand a client the keys without holding our breath.