← All insights

AI Security Risks: Closing the Gap betwen Innovation and Protection

20 November 2026 · David Patterson · AI

AI Security Risks: Hero Image
AI is accelerating the development and deployment of new technology. But are organisations adapting their security practices quickly enough to keep pace?

There is considerable pressure on businesses to adopt emerging technologies quickly. Whether it’s artificial intelligence, autonomous agents or new cloud services, the potential advantages are considerable. FOMO is real and nobody wants to discover that their competitors have gained a significant advantage while they’ve been carefully evaluating the risks.

Unfortunately, cybercriminals understand this pressure too, and the rush to adopt new technologies can introduce AI security risks that organisations haven’t fully anticipated.

Emerging technology often combines new capabilities with relatively immature development practices, unfamiliar dependencies and security arrangements that haven’t yet been tested extensively in real-world environments.

The technology itself isn’t necessarily insecure, it is just that the processes surrounding its adoption may not have caught up.


New technology, familiar vulnerabilities

A recent incident involving Microsoft’s Agent Framework provides a useful example.

In November 2026, two malicious versions of the framework’s Python package were published to the widely used PyPI software repository. The modified packages contained code designed to steal credentials, including cloud access tokens and development secrets.

The particularly concerning aspect was that the malicious code could execute when Python started, without a developer explicitly running the affected framework.

The compromised versions were subsequently removed, but developers who installed them were advised to treat potentially accessible credentials as exposed.

This wasn’t an AI model behaving unexpectedly or being persuaded to ignore its instructions. It was a conventional software supply-chain attack against a component used to build AI applications.

And that’s what makes it interesting.

As organisations race to develop AI-powered systems, they are adopting new frameworks, libraries, integrations and deployment tools. Each creates another dependency that needs to be understood, maintained and secured.

Attackers don’t necessarily need to invent new techniques. They can apply familiar ones to an expanding collection of technologies.


AI Security Risks: When the bottleneck moves

AI is making it possible to develop and deploy software more quickly. Tasks that previously required days of development can sometimes be completed in hours.

But security assurance doesn’t automatically become faster at the same rate.

Reviewing dependencies, testing permissions, understanding data flows and establishing appropriate governance still require attention and expertise.

In effect, the bottleneck moves.

The ability to build and deploy new capabilities increases, while the capacity to evaluate and secure them may struggle to keep pace.

This is particularly important with AI agents.

Unlike conventional applications that perform a relatively narrow set of predefined operations, agents may be given access to multiple systems, information sources and tools. Some can initiate actions, retrieve sensitive information or interact with other services.

A vulnerability in one component can therefore have consequences beyond the component itself.

An agent with access to customer records, email systems and business applications presents a very different risk from a standalone chatbot answering general questions.

Microsoft has published guidance on managing the risks of autonomous AI agents, including controlling permissions, securing software dependencies and maintaining human oversight.

The question isn’t simply whether the AI behaves appropriately. It’s whether the entire system has been designed to limit the damage if something goes wrong.

The widening security gap
A line chart to illustrate how technology tends to outrun security assurance over time.

The temptation to deploy first and secure later

None of this is unique to AI.

We’ve seen similar challenges with cloud computing, connected devices and other technologies that became widely adopted before organisations fully understood their operational and security implications.

The commercial pressure is familiar. A new capability becomes available, competitors begin experimenting, customers start asking for it, and organisations want to demonstrate progress.

Security review can then become something to complete before launch rather than something that shapes the architecture from the beginning.

But emerging technologies arguably make this approach more dangerous because their dependencies, capabilities and patterns of use can change so quickly.

A system that was appropriately secured when first deployed may become exposed through a software update, a new integration or an expansion of its permissions.

Security needs to evolve alongside the application, not simply be assessed once.


Quantum computing presents another challenge

Looking further ahead, quantum computing illustrates a different aspect of the same problem.

A sufficiently powerful quantum computer could undermine widely used public-key cryptographic methods, including RSA and elliptic-curve cryptography, which help secure digital communications and authenticate systems.

That doesn’t mean quantum computers will suddenly break every form of encryption or automatically discover software vulnerabilities. The threat is more specific, and the timing remains uncertain.

Nevertheless, the implications are significant.

Information encrypted today may need to remain confidential for decades. If attackers collect that information now, they may eventually be able to decrypt some of it using future quantum capabilities.

This is one reason why standards bodies, including the US National Institute of Standards and Technology, have already developed post-quantum cryptographic standards.

Organisations don’t need to wait for a practical cryptographically capable quantum computer before beginning to understand where vulnerable cryptography exists in their systems.

The important lesson is that security planning must account for technologies that may change the assumptions on which today’s systems depend.


Keeping security aligned with innovation

The answer isn’t to slow innovation unnecessarily. Emerging technologies can deliver substantial benefits, and excessive caution carries its own commercial costs.

The UK’s National Cyber Security Centre has published guidelines for secure AI system development, emphasising that security needs to be considered throughout design, development, deployment and ongoing operation.

Managing AI security risks needs to be part of how organisations adopt and develop technology, rather than an exercise carried out just before deployment.

That means understanding the components and suppliers on which systems depend, limiting access to sensitive information, testing how systems behave when something goes wrong and ensuring that security arrangements can adapt as technology changes.

For AI agents, it also means treating permissions, tool access and human oversight as architectural decisions rather than optional safeguards added afterwards.

For quantum computing, it means beginning to identify cryptographic dependencies and planning for a transition that may take years.

Neither challenge can be addressed by a single security product or a one-off assessment.

The greatest security risk from emerging technology isn’t always the technology itself. It’s the gap between how quickly we adopt it and how quickly we adapt the systems, processes and assumptions that keep it secure.

Closing that gap is becoming an increasingly important part of responsible technology leadership.

Talk to KnowNow Information

At KnowNow Information, we help organisations understand, plan and implement emerging technologies, including AI, with appropriate attention to security, data governance and long-term sustainability.

If you’re considering how to introduce new capabilities while maintaining confidence in your existing systems, we’d be happy to discuss your approach.

Talk to KnowNow Information