Scout Shield
A fixed-scope security review of your AI deployment: prompt-injection probing, tool-permission audit, data-egress mapping, and a prioritized fix list — before an incident writes the report for you.
What you get
A structured review of an AI system you already run — an assistant, an agent, a chatbot a vendor sold you. We probe it with direct and indirect prompt-injection cases mapped to the OWASP LLM Top 10, audit which tools it can call with which permissions, trace where your data actually flows on every turn, and check what stands between the model and an irreversible action. You get a findings report ranked by severity, with a concrete fix for each.
What makes it different
This is a builder's review, not a checklist. We run production AI systems ourselves — the injection rules, permission scopes, approval gates, and egress controls we test for are the ones running in our own stack. Findings come with the fix pattern, not just the vulnerability class.
What it covers
Scope is agreed up front; a typical review covers one production AI surface end to end.
- Direct and indirect prompt-injection probing, with transcripts
- Tool and permission audit: what the AI can read, write, call
- Data-egress map: what leaves, to which providers, under what terms
- Approval-gate and escalation review on high-impact actions
- Grounding check: can it be made to state falsehoods with confidence
- Prioritized findings with a fix pattern per item
When it makes sense
You shipped an AI surface fast and governance came second. A vendor's chatbot has access you've never inventoried. Compliance or a customer is asking questions you can't answer from records. Or you're about to launch and want the probing done by someone other than your users.
How we ship it
Fixed scope, typically 2 weeks: access and inventory in the first days, probing and audit through the middle, findings walkthrough with your team at the end. Fixes are yours to implement — or a follow-up engagement hardens the system with the same guardrail patterns we run in production.
Questions people actually ask
We didn't build our AI system — a vendor did. Can you still review it?
Yes, and that's a common case: a chatbot a vendor sold you, with access nobody ever inventoried. We probe it with direct and indirect prompt-injection cases mapped to the OWASP LLM Top 10 (the standard list of AI security risks), audit which tools it can call, and trace where your data actually flows on every turn.
Is the probing safe to run against a live system?
Scope is agreed up front — a typical review covers one production AI surface end to end, with access and inventory settled in the first days before any probing starts. The point is that the probing gets done by us, under agreement, instead of by your users or by an attacker.
What do we get at the end, and how long does it take?
A fixed-scope engagement, typically 2 weeks: access and inventory first, probing and audit through the middle, then a findings walkthrough with your team. The report ranks findings by severity, and each one comes with a concrete fix pattern — with transcripts of the successful injections so you can see exactly what happened.
Are we obligated to hire you for the fixes?
No. The fixes are yours to implement — the report is written so your team can act on it directly. If you want help, a follow-up engagement hardens the system with the same guardrail patterns we run in our own production stack, but the review stands alone either way.
Frequently paired with
Related products
A one-week engagement that maps your AI-readiness — what's in your data substrate, what's missing, what AI capabilities you could deploy today, and what would need to be built first.
We take 60 days of your production AI traces, build a real eval set, and hand it back with a working harness — so you can finally measure whether your AI is getting better or quietly worse.