01 / 06
More attacks, more often
The NCSC's May 2025 assessment judges that AI will almost certainly make parts of cyber attacks more effective and efficient, leading to more frequent and more intense threats.
Why AI is a cyber threat to your organisation
AI is making cyber attacks faster, larger in scale and more convincing. The UK's National Cyber Security Centre (NCSC) and its Five Eyes partners say leaders need to act now. Here is what is changing, how an AI-enabled attack unfolds and what to do about it.
Why is AI a cyber security threat?
AI helps attackers do what they already do, faster and at greater scale. The NCSC judges that AI will almost certainly continue to make cyber attacks more effective and efficient, increasing the frequency and intensity of threats to 2027. Most of the change comes from improving familiar techniques, not inventing new ones.
02 / 06
Phishing that reads like the real thing
Generative AI can write lures without the spelling and grammar mistakes that often give phishing away. The NCSC warned it would make it difficult for everyone to judge whether an email or password reset request is genuine. Phishing hit 38% of UK businesses in the past year, the most common attack by far.
03 / 06
Less skill needed
AI lowers the barrier for novice cyber criminals, hackers-for-hire and hacktivists, according to the NCSC. It expects AI-enabled attack tools to spread to a wider range of criminal and state-linked groups.
04 / 06
Days, not weeks, to patch
The NCSC says the time between a vulnerability being disclosed and exploited has shrunk to days, and AI will almost certainly reduce it further. In May 2026 it told organisations to prepare for a "vulnerability patch wave", as AI helps attackers exploit long-standing weaknesses in software at scale and at pace.
05 / 06
Attack tools are improving fast
In tests run by the UK's AI Security Institute, the best AI model in early 2026 completed nearly six times more steps of a simulated attack on a company network than the best model 18 months earlier. A full attempt costs around £65, and the NCSC advises defenders to assume some attackers already have capable AI tools.
06 / 06
A widening divide
The NCSC expects a digital divide between systems that keep pace with AI-enabled threats and a large proportion that are more vulnerable. Weak cyber security will almost certainly give attackers more opportunity to misuse AI.
What do the numbers say?
Two snapshots from UK government research: how often UK businesses are attacked, and how quickly AI attack tools are improving.
Nearly 6 times more attack steps in 18 months
6×Nearly six times more attack steps completed by the best AI model in early 2026 than by the best model 18 months earlier
£65The approximate cost of one full attempt at the simulated attack, at current pricing
Attack steps completed
<2
9.8
15.6
22
the best AI model 18 months earlier
the best model in early 2026 (average)
the same model, given more processing time (average)
its best single run
The NCSC advises defenders to assume that at least some attackers already have access to capable AI tools.
What does AI change about a phishing email?
The NCSC warns that generative AI can write lures without the spelling and grammar mistakes that often give phishing away. This illustration shows the difference. The names, company and details are made up.
Without AI
- From
- Acounts Departement <accounts-dept@example.com>
- Subject
- URGENT!! Change of bank detials
Dear Valued Costumer,
We are writting to inform you that our bank account have been change. Kindly make all future payment to the new account detail in attachement, efective immediatly.
Failure to do so may result in your account been suspended.
Best Regard,
Acount Team
Highlighted: the spelling and grammar mistakes that often give phishing away.
With AI
- From
- Alex Morgan <alex.morgan@example.com>
- Subject
- Updated bank details ahead of Friday's payment run
Hi Sam,
Thanks for your time on the call with Priya last week. As part of our year-end changes, we have moved our business banking to a new provider, so I wanted to give you plenty of notice before Friday's payment run.
Our updated details are on the attached letter, on our usual letterhead. Could you confirm once your system has been updated?
Kind regards,
Alex Morgan
Finance Manager, Example Supplies Ltd
Highlighted: tailored details an attacker can find or guess. With no obvious mistakes, the request itself is the warning sign.
What still worksCheck any change of bank details by calling the supplier on a number you already hold, never one from the email, and have a second person approve payments.
How could an AI-enabled attack on your organisation unfold?
Most AI-enabled attacks follow a familiar path. AI makes each stage faster or more convincing, but the controls that stop them are largely the same. Here is how each stage can change, and what still works.
01Research
With AI
Attackers can research your people, suppliers and systems faster, using public information. The NCSC judges that AI makes reconnaissance more effective, efficient and harder to detect.
What still works
Know what you expose to the internet, keep it to a minimum, and keep an accurate list of your systems.
02The lure
With AI
Phishing emails and messages are fluent and tailored, and deepfake audio or video can imitate a voice or face your staff trust.
What still works
Verify payment, bank detail and access requests through a trusted channel, and use phishing-resistant MFA.
03Getting in
With AI
Attackers find and exploit weaknesses faster. The time between a flaw being disclosed and exploited has shrunk to days, and the NCSC expects AI to shorten it further.
What still works
Patch fast, ideally automatically, starting with internet-facing systems, and replace technology that is out of support.
04Moving through
With AI
AI models are rapidly getting better at the steps of an attack inside a network. In UK government tests, the best model completed nearly six times more steps of a simulated attack than the best model 18 months earlier.
What still works
Strong identity and access controls, least privilege, secure configuration and comprehensive logging.
05The payoff
With AI
The goal is usually money or data: a fraudulent payment, ransomware or stolen information. The NCSC judges AI will make attacks more impactful because attackers can analyse stolen data faster.
What still works
Prepare for incidents before they happen, with tested backups and a response plan you have practised.
Why is your own use of AI a risk too?
The AI tools you adopt become part of your attack surface. The NCSC warns that AI systems can be manipulated into misusing the access you give them, and that some of these weaknesses may never be fully fixed.
01 / 04
Prompt injection
AI models do not reliably tell instructions apart from the content they read, so hidden instructions in an email, document or web page can steer what they do. The NCSC says this may never be totally mitigated, so limit what the AI can do: don't give an AI that reads external emails access to privileged tools.
02 / 04
AI agents with too much access
Agentic AI can use tools, access data and take actions with little human input. The NCSC advises starting with tightly bounded pilots, never giving an agent unrestricted access to sensitive data or critical systems, and keeping meaningful human oversight.
03 / 04
Adoption is outpacing controls
Around a third of UK businesses (31%) are using, adopting or considering AI. Of those, only around a quarter (24%) have cyber security practices in place to manage the risks, according to the government's Cyber Security Breaches Survey 2025/26.
04 / 04
A code of practice to follow
The UK's Code of Practice for the Cyber Security of AI, published in January 2025, sets 13 principles across secure design, development, deployment, maintenance and end of life. It covers AI security in staff training and securing your AI supply chain.
What should your organisation do about AI threats?
Get the basics right, faster, and make AI cyber risk a leadership responsibility. In June 2026 the Five Eyes cyber security agencies, including the NCSC, said that success will come from getting the basics right.
- Own the risk at board level. Understand your risk, readiness and accountability, and give your security leaders the authority and resources to act.
- Reduce your attack surface. Find, and cut down, the systems and services you expose to the internet.
- Patch faster. Apply updates as soon as possible, ideally automatically, starting with internet-facing systems. Replace, or bring back into support, any end-of-life technology.
- Strengthen identity and access. Use phishing-resistant MFA and give people and systems only the access they need.
- Verify, don't trust. Confirm payments, bank detail changes and access requests through a trusted channel, so a convincing fake is not enough.
- Adopt AI securely. Start with bounded pilots, give AI agents the minimum access for the shortest time, keep human oversight, and follow the Code of Practice for the Cyber Security of AI.
- Prepare for incidents. Keep tested backups and a response plan you have practised, before you need them.
Questions
Can AI carry out a cyber attack on its own?
Not from start to finish, yet. In tests published by the NCSC and the UK's AI Security Institute in March 2026, no public AI model completed a full simulated network attack unaided. But performance is improving fast, and the NCSC advises defenders to assume that at least some attackers already have access to capable AI tools.
Is this only a problem for large organisations?
No. The NCSC expects a digital divide between systems that keep pace with AI-enabled threats and a large proportion that are more vulnerable. Weak basics give attackers more opportunity to misuse AI, whatever the size of the organisation.
Should we stop using AI tools?
Not necessarily. The NCSC suggests asking whether AI is really needed for a task, starting with tightly bounded pilots and applying existing cyber hygiene from the start. If you cannot understand, monitor or contain an AI agent's actions, it is not ready to deploy.
Can defenders use AI too?
Yes. The NCSC says defenders must use the same capabilities to gain an advantage, but that AI won't compensate for weak security foundations: it will amplify both strengths and weaknesses.
How does AI change phishing and deepfake fraud?
It makes messages more fluent and tailored, and lets attackers imitate a familiar voice or face. Our guide to AI-enabled phishing and deepfake fraud covers the controls that still work.
Sources
- The near-term impact of AI on the cyber threatNational Cyber Security Centre (NCSC) · 24 January 2024
- Impact of AI on cyber threat from now to 2027National Cyber Security Centre (NCSC) · 7 May 2025
- The AI shift in cyber risk: why leaders must act nowFive Eyes cyber security agencies including the NCSC · 22 June 2026
- Why cyber defenders need to be ready for frontier AINCSC and AI Security Institute · 30 March 2026
- Preparing for a "vulnerability patch wave"National Cyber Security Centre (NCSC) · 1 May 2026
- Prompt injection is not SQL injection (it may be worse)National Cyber Security Centre (NCSC) · 8 December 2025
- Thinking carefully before adopting agentic AINational Cyber Security Centre (NCSC) · 15 May 2026
- Code of Practice for the Cyber Security of AIDepartment for Science, Innovation and Technology (DSIT) · 31 January 2025
- Cyber security breaches survey 2025/2026Department for Science, Innovation and Technology (DSIT) · 30 April 2026
Free cyber risk review
Find your cyber gaps before attackers do
A free, no-obligation review for organisations of 50 to 5,000 people, with plain-English findings for your security team, finance and the board.