AI security · offensive testing
We test the AI you already shipped. Agent loops, tool calls, MCP servers,
retrieval pipelines. Then we tell you where it breaks, in plain enough language that
your engineers can go fix it the same week.
No credentials. No exploit. No malware. Just text sitting in a document your agent was asked to read, and now every tool that agent can reach is in play. OWASP has kept prompt injection at number one in every edition of its LLM Top 10, and this is why.
Services
Some teams are still working out what to build. Some are mid-build and want another set of eyes on the architecture. Some already have agents in production and need to know what an attacker could do with them. We're happy to come in at any of those points.
Getting AI into production without a rewrite six months later. We help with model and platform choices, agent and tool design, evaluation, and the unglamorous operational plumbing (logging, cost ceilings, approval checkpoints) that usually decides whether a pilot survives real users.
Architecture review, threat modeling, and design work for teams building with LLMs and agents. We read the code, sit in your design reviews, and tell you which controls actually matter for the way your system is built. Not a generic checklist.
Hands-on adversarial testing against your models, your agents, and everything those agents can reach. We test what you actually deployed: connected to tools, wired into real data, running most of its decisions with nobody watching.
How an engagement runs
A call, an architecture diagram, and a list of what the agent can reach. We'll tell you what's worth testing and what isn't.
Hands on the system. You get findings as we confirm them, not all at once at the end.
A working session with the engineers who own the code, plus a report naming the request, the trace, and the fix.
Once you've shipped the fixes, we re-run the findings from the report. One pass, within 30 days of delivery, included in the price. Past 30 days it's billable, so it's worth booking early.
About us
NullTrace Security is a small, senior team. No pyramid, no bench of juniors learning on your environment. Whoever scopes your engagement is the same person who runs it and the same person sitting across from your engineers at the readout.
Between us we've done security architecture and detection engineering inside large engineering organizations, real offensive work across networks, applications, cloud, and red team, and a good amount of plain software engineering. That last one matters more than it sounds. We've shipped and maintained production code, so we know the difference between a finding that's satisfying to write up and a fix that's actually cheap to land.
It shows up in the reports. Findings arrive looking like engineering tickets: the exact request that triggered it, the trace, the blast radius, and a fix we would sign off on in a code review. Nobody on your team should have to translate our writeup before they can work from it.
We test AI the way it actually ships. Connected to tools, wired into real data, making most of its decisions without a human in the loop.
That matters, because a lot of AI security advice still treats the model as the whole system. It isn't. The interesting failures live in the seams: what the agent is allowed to call, whose identity it calls with, what it'll believe from a document somebody handed it, and how far one sentence of untrusted text can travel before anything stops it.
We keep up with the frameworks too, because clients ask. OWASP now maintains three separate lists that touch this work: the Top 10 for LLM Applications, the Top 10 for Agentic Applications, and the MCP Top 10, which is still in beta. We'll map findings to whichever of those you report against.
Get in touch
First conversation is a scoping call, not a sales call. If your system doesn't need testing yet, we'll say so.
Direct line
info@nulltracesec.comHelpful in the first email
