Series 02 · AI Security
Proactive and reactive. Before something gets through, and after.
An agent with tool access is a new attack surface, and the interesting failures are not the ones the frameworks warn about. Prompt injection arrives inside data you trusted. Credentials end up in a context window. An agent takes an action nobody authorised because nobody wrote down what it was allowed to do. The proactive half is finding those before someone else does. The reactive half is what happens the week after something got through — containment, then repair, then the change that stops the repeat.
What this looks like
Real problems, real shapes.
A few scenarios that sit squarely in the AI Security practice. Your problem probably rhymes with one of them.
01
You shipped an agent with tool access
It can read, write, and act. Nobody has written down what it is not allowed to do.
02
Something already got through
Contain it, repair it, and change the thing that let it happen. In that order.
03
A client is asking questions you cannot answer
Procurement wants to know how the agent is bounded. You need a real answer, not a policy document.
How we engage
Choose the shape that fits.
Three engagement models per practice. Every one of them is outcome-focused and written into a one-page agreement before work starts.
01 · Model
Security Audit
A review of what you have running, with findings ranked by what an attacker would reach first.
Pricing
$2,500
Best for
Something is live and you have not looked at it this way yet.
02 · Model
Remediation
Fixing what the audit found. Deliberately not fixed-price — remediation cannot be scoped before the audit, and a number quoted blind is a guess.
Pricing
From $7,500, scoped from audit
Best for
The audit found things and you want them closed.
03 · Model
Monitoring
Ongoing watch with findings surfaced when they matter, not a dashboard nobody opens.
Pricing
$750/mo
Best for
The surface keeps changing because you keep shipping.
Recent work
Scrlpets is the worked example.
A marketplace built, launched, and still being iterated in-house — which means the pipeline underneath it has been run against real users rather than described in a deck. The portfolio carries everything else, each entry with its actual state.
See the portfolioStart a project
Ready for AI Security?
Tell us what you're trying to build. We'll tell you what it'll take — or point you somewhere better if we're not the fit.