For most of the history of computing, finding a serious security flaw took a skilled human weeks or months of careful work. That is changing fast. In 2026, AI models have found bugs that human experts and automated testing tools missed for decades — and the same capability is now in the hands of attackers.
This isn’t a future problem. The PHP2Shell vulnerabilities that took government websites offline in October were found by researchers working with AI tools. Here is what has changed, and why it matters even if you never write a line of code.
What AI Can Do Now
The clearest example came in April, when Anthropic published an assessment of an unreleased model called Claude Mythos Preview. In testing it independently found and exploited flaws across major operating systems and web browsers, including:
- a 27-year-old bug in OpenBSD, an operating system known for its security, that let a remote attacker crash any machine running it;
- a 16-year-old flaw in FFmpeg, video software used almost everywhere, that had escaped years of automated testing;
- a 17-year-old flaw in FreeBSD (CVE-2026-4747) that it used to gain full control of a server without logging in.
In one browser test, the previous model produced two working attacks in several hundred attempts. Mythos Preview produced 181. Anthropic has not released the model publicly. Instead, through a programme called Project Glasswing, it is giving access to selected technology companies and open-source developers so they can fix flaws before similar tools are widely available.
Others are doing the same. OpenAI’s “Patch the Planet” programme gives security researchers AI tools to hunt for flaws in widely used software, and that is how the firm Calif found PHP2Shell. Google, Wiz and a growing list of start-ups have similar tools.
The Numbers Are Already Moving
Google’s Threat Intelligence Group published a detailed analysis in October that shows how quickly things are changing:
- From January to August 2026, Google recorded 141 vulnerabilities exploited in real attacks — more than the 127 recorded in all of 2025.
- The monthly average rose from about 10 in 2025 to 18 in 2026.
- Flaws found by AI tend to be more serious: half led to remote code execution, compared with about a quarter of flaws found by other means.
Google also pointed to a flaw in BeyondTrust remote-support software (CVE-2026-1731), found by an AI research tool, that was being exploited by attackers within four days of being announced.
Attackers Have the Same Tools
The uncomfortable truth is that a tool that can find a bug for a defender can find it for an attacker too. Google reported the first known case of a criminal group holding a zero-day exploit — an attack for a flaw nobody had fixed yet — that had been written with the help of generative AI. In a later report, it described criminals sharing and testing AI tools built to hunt for vulnerabilities.
AI is also very good at the less glamorous work: reading a security patch, working out what it fixed, and building an attack for everyone who hasn’t installed it yet. That’s why the gap between a fix being released and attacks starting keeps shrinking.
Australia’s own Annual Cyber Threat Report, released by the Australian Signals Directorate in October, makes the same point. In his foreword, Deputy Prime Minister Richard Marles described AI as accelerating the scale, speed and complexity of cyber threats. The report notes that most flaws are now exploited within two weeks, some within 48 hours of a patch.
What This Means for Your Business
You don’t need an AI model to defend yourself against AI-assisted attacks. You need to be quicker at the basics.
- Patch in days, not months. A monthly patching cycle used to be acceptable. It isn’t any more. We covered why in Microsoft Now Says Patch Within Three Days. Turn on automatic updates wherever you can, and have a plan for the systems that can’t update themselves.
- Know what you have. You can’t patch a system you’ve forgotten about. Keep a simple list of your servers, websites, firewalls and cloud services, and who looks after each one.
- Shrink what’s exposed. Every system on the open internet is something an AI-assisted attacker can scan. Put remote access behind a VPN or Zero Trust service and retire old websites and test servers.
- Replace software that is out of support. Old software doesn’t get fixes. As AI finds more old bugs, unsupported systems become easier targets every month.
- Plan for the patch that comes too late. Multi-factor authentication, limited admin rights and tested, offline backups mean one exploited flaw doesn’t become a business-ending incident. These are the core of the Essential Eight.
There Is Good News
The same technology is helping defenders. Old, hidden flaws are being found and fixed in software everyone relies on, and security tools are getting better at spotting attacks. In the long run, AI should make software safer. In the short term, though, there will be a wave of newly found flaws, and the businesses that patch quickly will be the ones that ride it out.
How Electriclatte Can Help
Our cybersecurity services include patch management, vulnerability scanning and Essential Eight assessments, sized for small and medium businesses. If you’d like to know how quickly your business could respond to the next critical flaw, get in touch.
Sources
- Google Threat Intelligence Group (R. Grunewald, S. Mazumdar, K. Vanderlee), Vulnerability Discovery and Exploitation Trends in the AI Era, 1 October 2026.
- Help Net Security, Anthropic’s new AI model finds and exploits zero-days across every major OS and browser, 8 April 2026.
- GIGAZINE, How Calif and Wiz found PHP2Shell, 22 October 2026.
- Australian Signals Directorate, Annual Cyber Threat Report 2025–26, 26 October 2026.
- The Cyber Express, ASD Cyber Threat Report: AI Speeds Up Attacks On Australia, 29 October 2026.
Photo by Steve A Johnson on Unsplash.