Guide · Cybersecurity & AI

AI in Cyber Attacks and Defence: What Is Actually Happening, and What It Changes for You

Attackers are using AI to find vulnerabilities and write exploits faster, and the first fully automated exploitation frameworks, with no human in the loop, were reported in November 2026. Defenders are using it to find the same bugs first. The published evidence on both sides, what it means for the attacks that reach your organisation through suppliers and websites, and what it does not change.

The short answer

AI has changed the speed and the cost of attacks, not their mechanism. The published evidence shows AI finding real vulnerabilities, writing working exploits, and in a small number of documented cases running most of an intrusion on its own. It also shows AI finding bugs before attackers do and patching them in minutes. What it has not changed is where attacks land: a phished credential, an unpatched system, a vendor script your site trusts. The controls that stop those are the same ones as before, and the honest consequence of AI is that you have less time to put them in place.

If the attack that worries you is one through your website's suppliers, start with the web supply chain guide and the free website script check. This guide is about what AI adds on both sides.

What attackers are doing with AI, with the evidence

Most of the public claims about AI powered attacks are vendor marketing or speculation. These are the cases with a primary source behind them.

What was reportedDetail and source
An intrusion campaign run mostly by AI, November 2025Anthropic reported that a group it assessed with high confidence to be Chinese state sponsored manipulated its Claude Code tool into attempting intrusions at roughly thirty organisations, succeeding in a small number of cases, with the AI performing 80 to 90 percent of the tactical work and humans choosing targets and approving each stage. The tasks were presented to the model as routine technical work so it did not see the whole chain. The tools were ordinary open source ones, scaled by orchestration, and the model made mistakes along the way. Anthropic, 13 November 2025.
Exploitation with no human in the loop, November 2026Google's Threat Intelligence Group reported finding an exposed server running a multi agent framework that compared web application code before and after patch announcements, drafted exploits and ran them, the first time it had seen the whole chain automated. Vulnerability disclosures and exploits nearly doubled between January and August 2026, and exploited zero days reached 22 in August against a 2025 monthly average of 8. Open source agentic penetration testing frameworks are being pointed at victims, and the FBI issued an alert on 13 November 2026 about one of them in use since August. Help Net Security on the GTIG report, 17 November 2026.
Exploit code at volume, much of it brokenThe same report notes that within days of two PHP patches on 21 October 2026, hundreds of non working proof of concept exploits appeared online. AI produces exploit attempts faster than it produces working ones, which raises the noise as well as the risk.
AI used to build malware, not to run itCheck Point's 2026 AI security report separates two patterns. In the common one, AI is used during development, so the finished malware shows no sign of it at run time and looks like any other. The rarer one calls a model while running. Check Point, 2026.
Stolen AI access as the prizeAttackers hijack cloud AI accounts to run their own workloads and resell the access, with fraudulent charges in some cases passing $100,000 in a day. An API key is now loot in the way a database credential has always been. Same GTIG report.

Two things are worth noticing about that table. The mechanism in every case is familiar: an unpatched vulnerability, a phished account, a trusted tool pointed at someone else's systems. And the one thing nobody has published is AI as the way into a web supply chain incident. The Optimizely compromise of 16 November 2026 has no published cause at the time of writing, and anyone who tells you AI did it is guessing. What AI does to that class of attack is make the next one cheaper to build and quicker to scale.

What defenders are doing with AI, with the evidence

What was reportedDetail and source
A bug found before it was exploited, July 2025Google reported that its Big Sleep agent found a memory corruption flaw in SQLite, CVE-2025-6965, which Google said was known only to threat actors and at imminent risk of exploitation. It was patched in version 3.50.2 before any attack. Google did not publish how it knew the flaw was about to be used. The Record, July 2025.
Automated finding and patching, August 2025In the final of DARPA's AI Cyber Challenge, seven teams' systems found 54 of 63 planted vulnerabilities across open source projects and patched 43 of them, and along the way found 18 real vulnerabilities and patched 11, with no human in the loop during the competition. DARPA, 8 August 2025.
An AI at the top of a bug bounty leaderboard, June 2025An autonomous penetration testing system reached first place on HackerOne's United States leaderboard after submitting around a thousand vulnerability reports in three months, with findings reviewed by people before submission. TechRepublic, June 2025.
Reading what a human would notThe unglamorous use, and the one most organisations can start tomorrow. A tag manager container with 300 tags, a Content Security Policy with 60 sources, a year of firewall logs: a model reads them in seconds and lists what should not be there. Our website script check is a small example of the same idea without a model, and the full read we do on a site uses one.
The honest shape of it

The same capability is on both sides, so the advantage goes to whoever applies it first to the boring work. An attacker's model is scanning for the unpatched system and the open script host. A defender's model can find the same thing on Monday, and the question is only whether anyone pointed it there. The organisations that are losing are not losing to cleverer attacks. They are losing the race to the basics.

What it changes for you

  • Patch windows shrink. A framework that reads a patch and writes the exploit removes the days you used to have between a disclosure and a working attack. Internet facing systems need patching in days, not quarters, and the ones you cannot patch need to be behind something that can be.
  • Phishing gets better and more personal. Fake recruiter profiles, fluent messages in any language, lures tuned to the person. The human perimeter was already the one that failed most often. Human risk is the real security perimeter.
  • Your suppliers are the attack surface. The web supply chain incidents of the last two years all went through a vendor, not through the victim. AI makes it cheaper to find the vendor with weak access controls and to write the payload once the attacker is in. The defence sits on your side: what the browser may run on your pages. The web supply chain guide.
  • Your AI accounts are now credentials worth stealing. Treat API keys and agent tool configurations as you treat database passwords: scoped, rotated, monitored for unusual volume. The GTIG report describes attackers editing a developer's AI tool configuration so that starting the tool installed a backdoor.
  • Noise rises. Hundreds of broken exploits and a flood of AI written phishing mean more alerts that are nothing and more that are not. Detection that depends on a person reading everything will not keep up. What managed detection and response does, and when you need it. Faster attacks make the speed of detection and response the deciding control; the MDR Review is a free independent read on yours.

What it does not change

Where attacks land. Every documented AI assisted intrusion still needed one of the old doors: a credential, an unpatched system, a trusted tool or script. Multi factor authentication, patching, least privilege, a Content Security Policy, backups that cannot be altered. None of these cares whether a person or a model is on the other end. The organisations that had them in place on 16 November 2026 were not in the 118 whose sites ran the Optimizely payload in full. The technology of the attacker changed. The reason it worked did not.

What this means for a board

Three questions. How long does it take us to patch an internet facing system once a fix is published, in days. Which suppliers' code runs on our website and in our build pipeline, and who approved each. And where are our AI accounts and keys, who holds them, and what would we see if one were used ten thousand times overnight. If the answers are not known, the attacker's model will find out before you do.

Talk to us about where you stand

What to do, in order

  1. Measure your patch time for internet facing systems and get it under a week. Put what cannot be patched behind something that can.
  2. Inventory your supply chain: the scripts on your website, the tags in the container, the packages your developers build with. The website script check covers the first two for one page.
  3. Set the controls that do not care who wrote the attack: a tight Content Security Policy, multi factor authentication everywhere, least privilege, immutable backups.
  4. Treat AI access as a credential. Scope keys, rotate them, alert on volume, and review the configuration files of AI tools your developers run.
  5. Use AI on the defending side for the reading work: containers, policies, logs, configurations. It is the cheapest win available and most organisations have not taken it.
  6. Train people for the better phishing, and assume some of it will land. Detection and response matter more when prevention is probabilistic.

How C4C helps

We are a cybersecurity consultancy with decades on the vendor side, so we know what the tools do and what the claims are worth, and we are an AI advisory practice as well, so we know what a model can and cannot do for a security team. We start with advice: an honest read of where you stand against the attacks that are actually happening, in writing, with the gaps ranked by the risk they carry rather than by what is easiest to sell. Where the plan needs tooling, such as detection and response, email security, a web application firewall or an AI platform for the reading work, we apply our IDEAL framework to identify the need, decide between the options on your numbers, execute the implementation, make sure it is adopted and manage it across its lifecycle. We select and supply the right software where that helps, and we tell you plainly when the gap is configuration you already own. The cyber security posture review is where most of these conversations start.

Talk to us about where you stand

Tell us what worries you most: patching, suppliers, phishing, AI accounts, or all of it. Within two working days we will come back with an honest view of where you are exposed and what to do first. Free, written, and no obligation.

Prefer email? Reach us directly at hello@c4cgroup.co.uk.

Frequently asked questions

Are AI powered cyber attacks real, or marketing?

Both exist. The documented cases are real: Anthropic reported in November 2025 that a state sponsored group ran an intrusion campaign against roughly thirty organisations with its model doing 80 to 90 percent of the tactical work, and Google's Threat Intelligence Group reported in November 2026 a framework that found and exploited vulnerabilities with no human in the loop. Most of what is sold as "AI powered attacks" in vendor material is ordinary attacks with a new label. The test is whether there is a primary source.

Did AI cause the Optimizely compromise?

Nobody knows, because Optimizely had not published how the attacker gained write access to customer projects at the time of writing. Report URI lists a compromised system, a compromised credential with cross customer access, or something else. Anyone attributing it to AI is guessing. What is true is that AI makes the next attack of that kind cheaper to build, and that the defence, a Content Security Policy that allows script from your own origin only, works whoever wrote the payload.

Can AI find vulnerabilities before attackers do?

Yes, with published results. Google's Big Sleep agent found a SQLite flaw in July 2025 that Google said was known only to threat actors, and it was patched before any attack. In DARPA's AI Cyber Challenge final in August 2025, automated systems found 54 of 63 planted vulnerabilities and 18 real ones in open source code, and patched most of them. An autonomous penetration tester topped HackerOne's United States leaderboard in June 2025. The capability is real on both sides, and the advantage goes to whoever applies it first.

What should we change first because of AI?

Patch speed. A framework that reads a patch and writes the exploit removes the days you used to have between a fix being published and an attack. Get internet facing systems patched in days, and put what cannot be patched behind something that can. After that, inventory the code that runs on your website and in your build pipeline, and treat AI accounts and keys as credentials worth stealing.

Should we use AI in our own security team?

Yes, and the best first use is the reading work nobody has time for: a tag manager container, a Content Security Policy, firewall and identity logs, cloud configurations. A model reads them in seconds and lists what should not be there. Keep a person deciding what to do about it. Detection and response platforms already use models for triage, and that is where the benefit is largest for a small team.

Are our AI accounts a security risk?

They are a credential and a target. Attackers hijack cloud AI access to run their own workloads and resell it, with fraudulent charges reported above $100,000 in a day, and have edited developers' AI tool configuration files so that starting the tool installed a backdoor. Scope and rotate keys, alert on unusual volume, and review the configuration of any AI tool your developers run.