For thirty years, ‘too hard to exploit’ was a security strategy. This month, AI retired it. If your encryption still runs on a general-purpose software stack, every layer of that stack is now a target that machines can find, weaponise and hit faster than any human can patch.

A fourteen-year-old flaw. Found in 2024. Ignored. Exploited by AI in weeks.

This week, governments on both sides of the Atlantic pulled websites offline. Thirty-two UK government services went dark and four more were partially unavailable. The National Cyber Security Centre fired out more than 50,000 warnings to British organisations in little over a day. Cloudflare counted tens of thousands of exploit attempts within 45 minutes of disclosure, hundreds of thousands within two hours, and millions by the evening.

The cause was not a new bug. It was a flaw that had sat in one of the most widely deployed pieces of software on the planet for fourteen years, underpinning roughly 70% of the web. Human researchers had actually found a version of it in 2024 and set it aside, judging it too complicated to turn into a working exploit. That judgement held for two years. Then frontier AI models looked at the same code and turned the ‘unexploitable’ into remote code execution.

“Frontier AI models substantially reduced that barrier, turning a longstanding flaw into an exploit… This changes the economics of both discovery and exploitation.” Wiz Research, disclosure write-up, October 2026

The researchers who found it described it as ranking among the most severe vulnerabilities ‘arguably in software history’. OpenAI, whose models were involved in the discovery, was candid about what it means for the other side: ‘To our knowledge, frontier AI found these vulnerabilities before attackers did… the same capabilities can also help attackers rapidly develop offensive tools and exploits’. OpenAI, October 2026

Read that twice. The defenders got there first, this time, by weeks. The capability that found the flaw is not proprietary to defenders. It is a general property of the models, and the models are everywhere.

This was not a one-off. It is the new baseline.

Go back six months and the pattern is already there. In April, Anthropic disclosed that its Mythos model had, inside a few weeks, found thousands of previously unknown vulnerabilities across major operating systems, browsers and open-source libraries. Among them: a 27-year-old bug in OpenBSD, a 16-year-old flaw in FFmpeg that had survived more than five million automated fuzz tests, and a 17-year-old remote code execution bug in FreeBSD that the model discovered and exploited autonomously from a single prompt. In a benchmark against the Firefox JavaScript engine it converted known vulnerabilities into working shell exploits 72% of the time. Anthropic’s own red team noted that none of this came from security-specific training; it was, in their words, ‘a downstream consequence of general improvements in code, reasoning, and autonomy.’

In September, Anthropic’s threat intelligence report described live criminal and state operations in which one stolen developer token was turned into full cloud administrative control in roughly three hours, in which a Chinese-speaking espionage group ran ‘agent swarms’ with persistent campaign memory against around 50 organisations, and in which a single vulnerability-research workflow produced more than a dozen candidate zero-days in a month.

In October, Microsoft’s Digital Defense Report put numbers on the compression. The median time from a vulnerability being discovered in the wild to being weaponised is now well under 24 hours. 2026 is on course for a record of roughly 72,000 tracked CVEs. In Microsoft’s own test environment, frontier models took control of an entire enterprise domain through a 32-step attack chain.

“AI is changing the physics of cybersecurity.” Microsoft Digital Defense Report 2026

And the NCSC, speaking on its tenth anniversary, was blunter still about who is using these tools. All four of the UK’s principal state adversaries, China, Russia, North Korea and Iran, now use AI to increase the output of their attacks.

“They don’t need sleep, coffee, food or anything like that.” Paul Chichester, Director of Operations, NCSC, October 2026

The warning from the top of government

Yesterday, the UK Defence Secretary stood in a munitions factory and told the country to stop using polite language for what is happening. Terms like ‘hybrid warfare’ and ‘grey zone’, he said, do not convey the reality to the public. He prefers ‘state-sponsored terrorism’. He asked whether businesses were prepared for ‘increasingly sophisticated cyberattacks’, for disruption to essential supply chains by hostile states, and then put the question in terms anyone can picture:

“How would we cope if the payment systems of the supermarkets went down?” Wes Streeting, UK Defence Secretary, 22 October 2026

These threats, he said, are ‘not hypothetical but real, and not somewhere else but increasingly here’, and ‘we have to step up and take seriously our national resilience as well as our national defence.’

He is not alone. In June, the head of the NCSC told the RUSI Annual Security Lecture that the Centre had handled more than 200 incidents against UK critical national infrastructure and its supply chain in a year, around three-quarters of them linked to state actors, and that the NCSC assesses attackers will likely be using AI-enabled capabilities to exploit known vulnerabilities in legacy technology at scale across CNI by 2028.

“The many vulnerabilities that organisations tolerate today will be exploited in conflict tomorrow… In cyberspace, we are not preparing for tomorrow’s conflicts, to some degree we are fighting them today.” Dr Richard Horne, CEO, NCSC, June 2026

Back in May the NCSC’s Chief Technology Officer warned every organisation to prepare for a coming ‘patch wave’ as AI accelerated vulnerability discovery. Five months later it arrived, and he was the one issuing the 50,000 warnings.

What this does to the attack surface

Strip away the headlines and three things have changed, permanently.

• Latent is now live. Decades of technical debt in operating systems, libraries, drivers and firmware used to be protected by obscurity and difficulty. AI has removed both. Every line of general-purpose code you run is a candidate.
• The patch window is gone. As Professor Alan Woodward put it this week, ‘there’s always a window between the flaw being publicly known about and it being patched.’ That window is now measured in minutes, and attackers are automated while most patching is not.
• Trust in credentials is collapsing. Stolen tokens and residential proxies make hostile access look legitimate, and AI agents can turn one credential into an entire estate before a SOC analyst finishes their coffee.

Now look at where most organisations keep their encryption. It sits at the top of a software stack: an application, calling a crypto library, running on an operating system, over drivers and firmware, on a CPU with shared memory. Every one of those layers is general-purpose code with a patch history. Every one of them is exposed. And the keys, the one thing the whole arrangement exists to protect, live in system memory, reachable through side channels and privilege escalation. When the OS falls, the keys fall with it.

That stack was an acceptable risk when exploits took skilled humans months to build. It is not an acceptable risk when they take an agent an afternoon.

Shrink what can be attacked. Move what cannot be stopped.

The alternative is not to patch harder. It is to give attackers less to attack. Sitehop puts the encryption path onto fixed-function logic in an FPGA inside a single hardware boundary. There is no general-purpose operating system in the data plane, so there is no OS vulnerability to find. Keys are generated, stored and used inside that boundary and never leave it, designed to FIPS 140-3 Level 3; there is no memory for an escalated process to read. Secure boot and hardware attestation make the device tamper-resistant by construction, and zero-trust segmentation is anchored to hardware identity rather than to a credential that could have been bought on a forum last Tuesday.

This is what shrinking the attack surface means in practice: fewer layers, fewer lines of code, fewer things that can be wrong, and no soft target in memory for an AI-written exploit to go after.

It also means moving at speed. Encryption in silicon runs at line rate, 1G to 100G, with no CPU tax and no ‘inspection’ hop where traffic is briefly in the clear. For defence, finance, critical infrastructure and OT, where links are contested, intermittent or denied, tunnels re-establish automatically when the link returns, without an operator in the loop. Security that waits for a human is security that loses to a machine.

Quantum is the next wave, and it is already breaking

Everything above is happening before the quantum threat matures. Adversaries are harvesting encrypted traffic today to decrypt it later, and the migration to post-quantum cryptography is a race with a hard deadline. A software stack that struggles to patch a 14-year-old memory bug in time is not going to re-platform its cryptography gracefully. Sitehop’s architecture is crypto-agile and PQC-ready: algorithms can be swapped and upgraded within the hardware boundary without rewriting applications or redesigning the network. Security evolves in step with the threat, not three years behind it.

The question for every board

The UK Defence Secretary asked the country a question about supermarket payment systems. Boards should ask themselves a sharper one: if an AI finds the next fourteen-year-old flaw in your software stack tomorrow, and the exploit is live within the hour, what is still protecting your data in motion?

If the honest answer is ‘the same operating system the exploit just took’, you do not have encryption. You have a promise, and the machines have stopped keeping it.

Shrink the attack surface. Put the keys where no software can reach them. Move at the speed of the threat with hardware encryption.

Book a Sitehop demo to see how hardware-enforced segregation and crypto-agile encryption keep a single breach from becoming a business-wide incident.

Stay in touch with Sitehop’s latest thinking, subscribe to our PQC Bulletin.

Or call us: +44 (0)114 478 2366

Sitehop. Engineered for resilience. Built for the life.

 

 

Sources

Wiz / OpenAI quotes, CVSS, timeline, 70% figure –https://securityboulevard.com/2026/10/most-of-the-worlds-websites-threatened-by-critical-php-flaws-found-with-ai/

14-year history, 2024 partial discovery, Cloudflare exploit counts, OpenAI ‘offensive tools’ quote –https://fortune.com/2026/10/21/ai-tools-discovered-the-php-vulnerability-that-could-take-down-70-of-the-internet-now-the-race-is-on-to-patch-it-before-ai-powered-hackers-attack/

32 UK services offline, 50,000 NCSC warnings (Whitehouse), Woodward quote –https://www.itv.com/news/2026-10-22/more-than-30-government-services-offline-amid-global-php-cybersecurity-threat

Baroness Lloyd and No.10 quotes –https://www.lbc.co.uk/article/74365c9975d1467496696a1be806a306-5HjdkGW_2/

Streeting speech, Telford, 22 Oct 2026 –https://newswav.com/article/prepare-for-state-sponsored-terrorism-and-cyber-attacks-to-cause-food-short-A2610_CxxTqf

Horne RUSI lecture: 200+ CNI incidents, 75% state-linked, 2028 assessment –https://www.ncsc.gov.uk/news/ncsc-ceo-hostile-states-linked-to-three-quarters-of-cyber-attacks

Chichester ‘no sleep’ quote; four adversaries using AI –https://www.resultsense.com/news/2026-10-15-ncsc-chichester-adversaries-ai-cyber-attacks/

NCSC ‘patch wave’ warning (Whitehouse, May 2026) –https://therecord.media/british-cyber-ai-patch-wave

Microsoft Digital Defense Report 2026 figures and quote –https://www.helpnetsecurity.com/2026/10/02/ai-cybersecurity-threats-microsoft-report/

Anthropic threat intelligence report, Sept 2026 –https://www.anthropic.com/threat-intelligence-report-september-2026

Mythos / Project Glasswing findings, April 2026 –https://oasis.security/blog/claude-mythos-project-glasswing-security-teams