With AI, cyberattacks come fast; it’s time firms patch faster

SINGAPORE - News on April 7 that Anthropic withheld the public launch of its latest artificial intelligence (AI) model, sharing it with only a tightly controlled group, has triggered a harsh reality check.

The San Francisco-based research firm said its latest Claude Mythos Preview is shared with only about 50 tech firms, including Microsoft, Google, Amazon and Apple (via an initiative dubbed Project Glasswing), as the tool could attack existing software without human action.

In a grave note on its website, Anthropic said: “Given the rate of AI progress, it will not be long before such capabilities proliferate, potentially beyond actors who are committed to deploying them safely. The fallout – for economies, public safety and national security – could be severe.”

It is no secret that AI models have been assisting hackers in finding software flaws for some years now, quickening the discovery of zero-day vulnerabilities from months to hours. Zero-day vulnerabilities are flaws unknown to the software maker, and thus, they have no fix.

But it was not until Claude Mythos Preview that AI models could also act autonomously – that is, they can find new vulnerabilities and generate code to exploit the flaws. With this tool, the barrier to connecting multiple minor vulnerabilities into a full system compromise has been substantially lowered.

That AI models can act without human intervention is an emerging reality that has happened sooner than most expected and is causing widespread alarm.

Following the April 7 news break, US Treasury Secretary Scott Bessent and Federal Reserve chairman Jerome Powell called a surprise meeting with Wall Street leaders over concerns that Claude Mythos Preview would usher in an era of greater cyber risk.

South Korea’s Ministry of Science and ICT reportedly held an emergency briefing with major domestic cybersecurity companies to discuss countermeasures. 

The Straits Times understands that the Cyber Security Agency of Singapore (CSA) has flagged the heightened risks to critical services operators in Singapore, including those running the country’s banking, energy and telecommunications systems.

In a statement to ST, the Monetary Authority of Singapore (MAS) said: “Financial institutions need to redouble efforts to strengthen their security defences, proactively identify and close vulnerabilities, and raise vigilance on cyber hygiene, including timely security patching. MAS is coordinating closely with the CSA to further strengthen support to critical infrastructure operators.”

CSA has also issued an advisory urging all firms to plan ahead to guard against such risks, even though these advanced AI capabilities have yet to be abused at this point.

The fears are real. 

So far, Claude Mythos Preview has discovered thousands of high-severity zero-day vulnerabilities, including in every major operating system and web browser, that have been missed for decades.

For instance, Mythos discovered a 27-year-old bug in OpenBSD, an open-source operating system based on the Berkeley Software Distribution model that is highly regarded for its uncompromising focus on security. The bug allowed for remote crashes.

This discovery is a jarring wake-up call as OpenBSD is used in scenarios requiring strong network security (firewalls, virtual private network gateways and intrusion-detection systems) and in critical services sectors such as finance, telecommunications, government and defence.

Experts may argue that because the vulnerability lies deep within the network, AI security tools would have already automatically detected and stopped attackers at the edge of the network to mitigate the high-impact risk.

But attackers do not need a constant stream of zero-day vulnerabilities, including critical ones, to succeed. Historically, three-quarters of all compromises involve one or more of just 10 known vulnerabilities, all of which have had patches available to organisations.

Case in point: The global spread of the WannaCry ransomware in 2017 that crippled factories, healthcare institutions and railway operations around the world exploited a known vulnerability called EternalBlue in Microsoft Windows’ Server Message Block protocol. A security patch had already been released two months before the attack. But more than 230,000 computers across 150 countries, including those belonging to the British National Health Service and FedEx, had not applied the fix.

In another high-profile example, Florida-based background check company National Public Data (NPD) in 2024 discovered a massive data breach involving 2.9 billion people in the US, Britain and Canada, including their social security numbers and physical addresses.

Hackers exploited unpatched vulnerabilities in NPD’s open-source Apache servers, for which a fix had been released two years before the attack. The stolen data was reportedly hawked for US$3.5 million (S$4.5 million) on the dark web. NPD’s original owner, Jerico Pictures, filed for bankruptcy and closed the site in late 2024.

In Singapore, a major data breach discovered in 2018 involving the personal data of 1.5 million SingHealth patients was also largely enabled by an unpatched server and other security lapses.

The serious incidents underscore a critical, longstanding issue that patches are not being applied fast enough, if at all, within most organisations. 

Unpatched software flaws – in other words, a fix was available but not applied – have consistently been found to be a primary enabler of cyberattacks and data breaches. 

Many organisations struggle to patch their systems as IT departments dread the task. 

Tech personnel often operate in a thankless, high-pressure environment where they are overlooked when systems run perfectly, but blamed immediately for any disruption. The high demand for continuous digital service availability often results in an “if it ain’t broke, don’t fix it” approach to patching to avoid risking operational disruption. 

Because tech systems are so complex, it often takes months to test patches to make sure they do not break any workflow or digital services. Imagine not being able to send funds, transact online or connect to mobile networks, all of which are deemed critical services in digital economies. 

Patch fatigue has also set in. Thousands of vulnerabilities are reported annually, with vendors issuing frequent updates. This essentially means applying at least three fixes a day.

Organisations frequently overlook the need to spend on automating patch management, ranking it below core activities like keeping systems running and software development.

As organisations digest the implications of Claude Mythos Preview, their top priority must be to invest in AI-powered patch testing, automating one of the most resource-intensive security tasks.

While it may not be possible to outpace flaw discovery, properly recognising the importance of patching is a solid start. 

What role does an individual play in this threat environment?

While the key lies in what organisations do to protect their customers, consumers can also take immediate steps, such as using different passwords for their different online accounts, enabling multi-factor authentication, and updating the operating systems of their devices promptly.

One should also be wary of clicking on links, which are frequent ruses to trick users into divulging confidential information.

If Claude Mythos Preview was a public relations exercise, it worked. Almost every reputable news outlet echoed Anthropic’s narrative. 

The media feeding frenzy that ensued also sparked a race for thought leadership in AI cybersecurity.

Anthropic’s rival OpenAI was reportedly working on an advanced AI model that would be released to a select group of partners rather than to the general public. Many cybersecurity vendors swarmed the opportunity with pitches for their own AI cybertools.

At the very least, this piece of news acts as a timely wake-up call for all organisations to rigorously review their security processes, particularly the automation of patch testing.

As Mr Ross McKerchar, chief information security officer at cybersecurity firm Sophos, said: “We can’t control the pace of AI-driven vulnerability discovery, but we can control how fast we respond, how we build products, how transparently we operate, and whether we treat patching as a feature or an afterthought.”