Every business rushing to deploy AI is running the same experiment: giving a language model access to customer conversations, internal data, or the ability to take actions, and hoping it behaves. Most of the time it does. The problem is the times it doesn't are rarely obvious until someone finds them, and by then they've usually found them by exploiting your customers, not by filing a bug report.
What actually goes wrong
These aren't theoretical risks. They're the everyday failure modes of AI systems deployed without anyone stress-testing them first:
- Prompt injection: a customer (or a document, or a webpage the AI reads) types something that overrides the system's instructions and makes it do something else entirely.
- Data leakage: a support chatbot tricked into revealing other customers' details, internal pricing, or the system prompt itself.
- Jailbreaks: guardrails around tone, topics or refusals bypassed with a cleverly worded request.
- Unsafe tool use: an agent given access to send emails, update records or issue refunds, with no check on whether it's being manipulated into misusing that access.
An AI system with access to your data is a new part of your attack surface, whether you built it or bought it.
How the audit works
We treat your AI system the way an attacker would, not by reading its documentation, but by actually trying to break it.
- Adversarial testing: structured attempts at prompt injection, jailbreaks and boundary-pushing across the system's real inputs.
- Access review: what data and tools the AI can reach, and whether that access is broader than the task requires.
- Compliance check: basic data-handling gaps that matter for UK businesses, especially where customer or employee data is involved.
- Written findings: each issue rated by severity, with a plain-English explanation and a fix.
From £1,500 for one chatbot or agent. Scoped to what you're running. Whether it's one customer-facing chatbot or a set of internal agents with system access, we quote a fixed scope before starting.
Who this is for
Firms that have already shipped an AI system (their own, or a vendor's) and want to know it's safe before it's their name in the incident report. That includes agencies offering AI features to their own clients, who need to be able to say the system has actually been tested. It's also for anyone about to sign off on a vendor AI tool and wants an independent check before the contract, not after the breach.
Frequently asked questions
What is an AI security audit?
A hands-on test of your AI systems (chatbots, agents, or vendor tools) to find where they can be tricked into leaking data, ignoring their instructions, or taking actions they shouldn't. You get a written report of what's exploitable and how to fix it.
What kinds of risks do you test for?
Prompt injection, data leakage through the model's responses, jailbreaks that bypass guardrails, and unsafe tool or system access given to an agent. We also check for basic compliance gaps around data handling.
Do you only audit systems you built?
No. We audit any AI system, your own custom builds or third-party vendor tools you've deployed. If it talks to customers or has access to your data or systems, it's in scope.
How is this priced?
From £1,500 for one chatbot or agent, scoped to the number of systems and access points involved. Get in touch and we'll give you a fixed quote before any work starts.
Related services
- AI Agents: support, research and data-gathering handled end to end, with a person on the decisions that matter.
- In-Product Copilot: a natural-language assistant built into your own product.
- AI Opportunity Audit: a fixed-price, ranked and costed map of where AI saves your firm hours. From £750.
- Worked example: the full audit deliverable for an 18-person recruitment agency.
The weekly note. One short email a week from Akashuin, our research blog: what UK firms are actually doing with AI, the numbers behind it, and one thing you can try that week. Free, and one click to unsubscribe.