Know what to fix in your AI codebase.
A human-reviewed audit of your agents, tools, prompts and model calls. Get findings tied to file:line, the reasoning behind each one, and a clear account of what we could not evaluate.
Audit Express · USD 2,500 · One repository, report and walkthrough.
Request an auditFor Python and TypeScript/JavaScript teams. Start with a 30-minute scope call; repository access is agreed afterwards.
Compare audit packagesA report your engineers can work from.
The differentiator is structural honesty. A zero from us means "we looked and there is nothing", never "we didn't look".
Findings anchored to file:line, each with a named reasoning path and a severity that respects code role. Example and vendored code is flagged as signal, not noise.
What we checked — and what we did not: which rule families evaluated your repository, which report
NOT EVALUATED, and why.A blast-radius review (Audit Profundo): the privilege surface of your repository, agents to tools to resources, with the controlled-shell discriminator.
An honest-limits letter in every delivery: what a static scan of a repository can and cannot see, and this detector's registered boundaries.
A walkthrough with your engineers: what is real, what is LOW by design, what one line of declarative config fixes.
Start with one repository. Go deeper when you need to.
One repository, full deterministic run, human triage, report and a 60 to 90 minute walkthrough.
Start here if you want to see what it finds in yours.
Everything in Express, plus the blast-radius review, a declarative-config session with your team, and a 30-day re-audit on your fixes.
The re-audit is what makes your fixes measurable, not just promised.
An audit per release or per month, a diff against the previous audit, and a short asynchronous walkthrough.
For teams that want governance to move with the code.
From scope call to delivery in one to two weeks.
The run itself is fast. The value is in the pin, the triage and the conversation.
Scope call, 30 minutes. Repositories, size, what worries you, which package.
Read-only access at a pinned commit. A collaborator invite or a tarball. The pin is recorded, so the audit is reproducible against that exact SHA.
Run and human triage, 1 to 3 business days. Every finding is read by a person before it reaches you.
Delivery and walkthrough. For Audit Profundo, the 30-day re-audit closes the loop on your fixes.
Three situations where a measured audit changes the conversation.
You sell to regulated buyers.
Banks, healthcare, defense and the public sector now ask how your agents are governed. A third-party audit with published precision is an artifact their procurement can file. A slide is not.
Your agents hold real privilege.
A browser session, a shell, a shared memory layer, a payments API. The blast radius of a steered agent is the privilege it holds. We map agents to tools to resources and tell you where a hijacked instruction lands.
You ship weekly with a small security team.
Model calls whose output nothing validates. Prompts that quietly encode a pricing rule. Documents that can talk to your agent as if they were you. Ordinary code review does not catch these. Ours is built for them.
We do not ask you to take our detection on faith. The homework is public.
Precision and recall per rule, over hand-labeled findings from real open-source products pinned to specific commits.
The false positives we found and eliminated are listed, dated, with the record that fixed them. So are the doctrines we measured and refused.
The page cannot disagree with the labels; the build would fail. Recompute it with one script, no dependencies.
Under NDA: the evaluator bundle, for the client who wants to audit the auditor.
Want to see it on code first? Ask your AI agent to run the free preview on any public repository — MCP endpoint mcp.divergia.ai/mcp, how it works. The audit itself never runs as a hosted service.
What the engine is measured on
Read the evidence- hand-labeled findings
- 3,823
- real open-source products, pinned to commits
- 20
- detection rules, precision and recall published per rule
- 20
- detection doctrines we measured and refused
- 5
Every number is computed from the labels at build time. Clone the evidence repository and recompute the page yourself:
git clone https://github.com/reoxb/probanza-evidence cd probanza-evidence python3 verify.py
Your code, our machine, your terms.
Who sees our code?
The people doing the audit at Divergia, and nobody else. The engine runs on our machine, not in your infrastructure and not as a hosted service. We read your repository once, at a commit you pin, from a read-only invite or a tarball you hand us.
Do you keep our repository?
We work from a read-only copy and record the pinned commit so the audit is reproducible. State your retention and access requirements at the scope call and we put them in writing before any access is granted.
Will you sign an NDA first?
Yes. If your process requires an NDA before access, we sign it before the scope call ends.
Does anything about our audit become public?
No. Client audits are private. The public evidence page measures the engine on open-source products only; your repository and your findings never appear there.
Is this a penetration test?
No. It is static analysis of your codebase: it finds exposure before it ships, it does not attempt attacks, and it declares what it did not evaluate. It complements a pentest; it does not replace one.
What if it finds nothing?
You still receive the applicability envelope: which rule families evaluated your repository and which reported NOT EVALUATED, and why. A zero from us means we looked. The price is the same.
Which codebases can you audit?
Python and TypeScript/JavaScript codebases, including agent frameworks and MCP tool manifests. Other languages are not supported today, and we tell you so before you pay.
A repository scan sees the code-shaped half of your risk. A default password on a live production account is a deployment fact, invisible to any static analysis, ours included. This is not a pentest and does not replace one. Our detector's registered boundaries are declared in the delivery, not papered over.
Tell us what you want reviewed.
Email us a short description of your product, its language and the risk you want to understand. We will use the scope call to agree what to audit and which package fits. No code or credentials needed in the first message.