You might have seen the headlines about AI finding critical security vulnerabilities in PHP, the language which powers the majority of websites online – including all WordPress websites.
Several major Australian government websites including the Australian Bureau of Statistics and the cyber.gov.au portal took down their websites as a precaution while updates were hastily arranged.
This approach is one way to avoid a new exploit in the wild, but it emphasises the need for speed and agility when it comes to new threats.
The hosting providers Mogul uses and recommends for clients – Pressable and WP Engine – began rolling out the security updates as soon as they were available. This fast response is a key advantage of choosing good managed hosting over cheap hosting, or having to scramble with DIY solutions.
A gathering storm
This security scare has been coming for a while now, and we fully expect to see more of these sorts of incidents as AI tools get better.
It’s not a fundamental issue with PHP or WordPress, or even with open-source software in general – proprietary software has exactly the same kinds of issues, often with fewer eyes working on them.
AI is really good at code, and newer models are proving exceptionally good at finding bugs and security vulnerabilities in code. As Anthropic (the creators of Claude) put it:
AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities.
Make it stop!
In theory AI models have security ‘guardrails’ built in to try to stop people misusing them or turning them to malicious purposes – a process known as jailbreaking.
However, it’s proving extremely difficult to completely prevent this, and some researchers think it may be fundamentally impossible, akin to the idea of ‘make no mistakes’ or ‘fix all the bugs’.
In response to the United States Government’s extraordinary order to restrict access to the latest Anthropic models in June, the company said:
We suspect that perfect jailbreak resistance is not currently possible for any model provider. Every safeguard used in the industry is vulnerable to non-universal jailbreaks (which can elicit some cyber information in specific circumstances), and it is likely that universal jailbreaks will eventually be found in the future.
Early models could be tricked into jailbreaking by simply asking them to ‘roleplay’ as a malicious AI, or to demonstrate the type of code a malicious person might want to generate.
The guardrails have become much more sophisticated now, but this battle is nowhere near won. And guess what is exceptionally good at finding sophisticated AI jailbreak techniques…
One silver lining
Powerful AI mixing with cybersecurity is not all terrible news. It can also be used by the good guys to find exploits and fix them.
In fact it turns out to be amazing at finding incredibly obscure vulnerabilities which have escaped notice in software for decades. Claude Mythos – one of the models temporarily restricted by the US Government as mentioned earlier – proved this earlier this year:
It found a twenty-seven-year-old vulnerability in OpenBSD that allowed an attacker to remotely crash any machine running the operating system simply by connecting to it. It found a sixteen-year-old vulnerability in FFmpeg’s H.264 codec that automated testing tools had hit the relevant line of code five million times without catching. It found and chained multiple vulnerabilities in the Linux kernel to achieve complete privilege escalation from an ordinary user account to full root control.
In a case of almost too much of a good thing, open source projects are seeing their bug report queues go through the roof as new vulnerabilities are found and reported.
It’s likely that this will all eventually reach a sort of equilibrium where the human-era software vulnerabilities are fixed, and new ones become rarer and more difficult to execute.
However in the meantime there’s a good chance it’s going to get worse before it gets better.
So what can you do to protect your digital presence and your customers?
The first and best thing you can do – and something we’ve recommended for years – is to stay on top of maintenance and software updates. While it may not stop ‘zero-day’ exploits which nobody has released a fix for, the fact is many websites are sitting online with known, published vulnerabilities, even though the fixes are only an update away – and have been for weeks or months or years.
The next most important thing you need is a plan for what happens if (or maybe when) you get attacked. You need good, reliable backups, and a recovery plan you can put into action to get back online as quickly as possible. That includes being able to figure out what actually happened, so you don’t immediately get hacked again.
These are the most basic steps every organisation should take, and doubly important if you are hosting customer data or personally-identifying information (PII).
There’s a lot more you can do to reduce your risks and improve your resilience, both at a technical level and at the organisation level – starting with a good digital partner who has your back.
Want to talk resilience? Get in touch
