By Nik Kairinos, CEO and Co-Founder of RAIDS (Rogue AI Detection System) and Brett Stonefield, Executive Director and Co-Founder

Investors are pouring money into AI at a historic pace. Few of them can say what their portfolio companies’ models are actually doing.

Investment in AI has grown at an extraordinary rate over the past few years, as venture capital (VC) investors race to back the company that will play a pivotal (and profitable) role in what many expect to be the defining technology of our time. In the UK alone, AI startups raised $12.6 billion in the first half of 2026, more than four times the total for the same period last year, according to HSBC.

This is not another article about AI’s limitless potential. It is about a blind spot in the due diligence behind all that capital, and a risk that many investors have yet to price in.

Here is a simple test. Ask an AI portfolio company to produce a record of how its models have behaved in production over the past 90 days. I suspect very few could give a satisfactory answer. And with so-called “rogue AI” incidents increasingly making headlines, many of them at the world’s largest AI developers, investors should be asking how exposed they would be if one of their hottest prospects became the next one.

Even the frontier labs are struggling

By now, most people in the industry will have heard about the recent incidents involving OpenAI and Anthropic. In July, OpenAI disclosed that models under cybersecurity evaluation had broken out of their test environment and reached Hugging Face’s production systems. Anthropic reviewed its own testing in response and found that its models had also taken unauthorised actions.

The UK’s AI Security Institute subsequently identified further cases of agents operating beyond their intended boundaries, raising fresh questions about how AI is deployed. These incidents involved pre-release models under test conditions, often with safeguards deliberately reduced. But they still show how increasingly autonomous systems can pursue their objectives through actions their developers never intended.

OpenAI and Anthropic have vast resources at their disposal, including dedicated research teams working to understand and contain exactly these risks. If the companies at the frontier of AI development are running into these problems, there is no reason to assume a startup will fare any better, however thoroughly it tests.

And detection may be the harder problem. Anthropic’s incidents only came to light because the Hugging Face incident prompted it to look. Rogue behaviour could be occurring on any internet-connected AI platform, and no amount of pre-launch testing will catch behaviour that only begins once a product is live.

For investors, then, a portfolio company that reports no incidents should not be given a clean bill of health unless there is evidence that its systems are actually being monitored. Absence of evidence is not evidence of absence, and where rogue AI is concerned, no news is not necessarily good news.

The financial consequences reach investors too

A rogue AI incident can do serious damage to the company responsible, particularly if it leads to a security breach or harms customers. And that exposure doesn’t stop at the company’s door: it extends to every investor whose returns depend on the business continuing to grow.

Reputational damage is hard to undo. A company once regarded as a leading prospect can quickly become untouchable if an incident calls the reliability of its technology into question. Not every incident can be prevented, but a company that monitors its AI continuously and keeps a reliable record of its behaviour is far better placed to respond quickly and contain the damage.

The same question resurfaces at exit. When investors come to sell their stake, a prospective acquirer’s due diligence may well ask for evidence of how the company’s AI has behaved in production, particularly if it sells into regulated industries. If the company has no reliable record, the buyer is being asked to absorb a risk it has no way of sizing. You don’t need to work in venture capital to know what that does to a valuation.

Commercial pressure is building, too. As more UK businesses move from experimenting with AI to deploying it across their operations, enterprise procurement teams are asking tougher questions about governance. Developers that can show their products are independently monitored will have answers ready. Those that can’t risk losing the deal.

Monitoring belongs in due diligence

The limitation of traditional testing and periodic audits is that they only show how a system performed under particular conditions at a particular moment. The real world is far less predictable. Once deployed, models encounter new inputs and operating environments, and agents may begin taking actions their developers never anticipated.

Continuous, independent monitoring makes it possible to observe those systems in production. It establishes a baseline of normal behaviour, flags any deviation from it, and creates a record of what happened before, during and after an incident.

Monitoring will not prevent incidents. If recent months are any guide, some are close to inevitable. But it is far better to know the moment unexpected behaviour begins than to discover it weeks later, once the damage has compounded. At that point, knowing where to look, and having evidence of what happened, is invaluable.

Building monitoring into a product during development is also far easier than retrofitting it once the business has scaled. And regulation is approaching: UK companies selling AI products into Europe will need to comply with the EU AI Act, whose high-risk obligations apply from December 2027 and August 2028.

Without venture capital, startup ecosystems around the world could not bring new AI technology to market. But if portfolio companies are to scale at the pace investors are looking for, they will need to prove they can account for what their AI is doing, with a clear audit trail to back it up.

So, the next time a VC is shown an impressive AI demo and ambitious growth projections, they should ask to see something far less glamorous: a reliable record of what the model has actually been doing since it went live. If the company can’t produce one, that is a gap in the investment case, and one investors would do well to uncover before they sign the cheque.