AI Continua / Security Audits
An AI security audit that ends with fixes, not a slide deck
We test your AI systems the way an attacker would, review how they are built, and hand your engineers reproducible findings ranked by real-world risk.
What we examine
Four workstreams, one combined picture of risk
An AI system can be robust in the model and fragile everywhere around it. We look at all of it.
Adversarial testing
Prompt injection (direct and indirect), jailbreaks, system-prompt extraction, data exfiltration, tool misuse and multi-turn manipulation, run against your actual deployment.
Architecture and data-flow review
Trust boundaries, authentication, retrieval permissions, secrets handling, logging, and what data reaches which model, vendor and region.
Agent and tool permission review
Which actions each agent can take, with what credentials, under what approvals. We test whether those limits hold when instructions are manipulated.
Governance and compliance mapping
Policies, inventory, ownership, incident response and documentation, compared with NIST AI RMF, ISO/IEC 42001, the EU AI Act and your own control framework.
How it runs
A clear sequence, with no surprises
Scope and rules of engagement
Agree on systems, environments, test accounts, data handling and stop conditions in writing.
Test and review
Automated attack suites plus manual testing, alongside design and configuration review with your engineers.
Report and walkthrough
Findings with evidence, severity and fixes. A live session so your team can ask questions and challenge ratings.
Retest
We verify each fix and mark findings closed, with regression tests left behind in your pipeline.
The report
Every finding can be reproduced and fixed
You get the exact input, the observed output, why it matters in your business context, and the change we recommend. Engineers do not need to guess what we meant.
- Executive summary written for the CISO and model owner, in plain language.
- Technical findings with severity, evidence, affected components and framework mapping.
- Remediation plan ordered by risk reduction per unit of effort.
- Residual risk statement that is honest about what cannot be fully closed.
Indirect injection reaches refund tool
Instructions hidden in an uploaded PDF caused the agent to issue a refund call without approval.
Retrieval ignores user role
A sales user could retrieve passages from HR policy documents through a paraphrased query.
System prompt extractable
Multi-turn role-play revealed internal routing rules and an internal URL.
Verbose errors in chat
Stack traces from the tool layer were shown to end users.
Engagement options
Sized to what you need to know
Scope and timing are confirmed after a scoping call, because they depend on how many systems you have and how they connect.
Focused assessment
One AI system
A fast, deep look at a single chatbot, copilot or agent, usually before launch or after a major change.
- Adversarial testing
- Configuration and design review
- Findings report and walkthrough
Full audit with retest
One system or a connected group
The complete four-workstream audit, including governance mapping and verification of your fixes.
- All four workstreams
- Framework mapping
- Retest and closure letter
- Regression tests for your CI
Continuous assurance
Ongoing
Scheduled re-testing and monitoring so your assurance keeps pace with model, prompt and data changes.
- Attack replays on each release
- Drift and abuse monitoring
- Regular reporting to owners
Audit questions
Is this a certification?
No. An AI Continua audit is an independent technical assessment mapped to recognized frameworks. It is not a formal certification under ISO/IEC 42001 or any other scheme, and it is not legal advice.
What access do you need?
Typically a staging environment, test accounts at several privilege levels, architecture documentation and time with the engineers who built the system. We agree on everything before testing starts.
How do you handle our data?
We work under NDA, limit access to what the scope requires, avoid real customer data where possible, and delete test data at the end of the engagement per the agreed terms.
Can the audit be shared with our customers or regulators?
Yes. We can provide a summary suitable for external sharing, separate from the detailed technical findings.
Do you also fix what you find?
If you want us to. Because our engineers also build agents and guardrails, we can implement the fixes ourselves or support your team while they do.
Know your AI risk before someone else finds it
Send us a short description of your AI systems. We will respond with a proposed scope.