Firms are racing to protect themselves from AI cyber threats
The article recaps a stretch some analysts have called “Black October” or “the Hackening”: AI helped discover a critical flaw in PHP, which underpins most websites, and the vulnerability was leaked before patches could be deployed; hackers then stole some $6.3bn from Binance; in November supermarket freezers across Europe had their temperatures cranked up, destroying millions of euros of food, and funds were stolen from French bank Crédit Mutuel Arkéa in a burst of fraudulent transactions. Governments were hit too (air-raid alert apps in Ukraine and Russia; data stolen from Denmark’s population registry, America’s Department of Homeland Security and the FBI; Japan declared a cyber-security state of emergency). On 3 November Richard Horne of Britain’s National Cyber Security Centre warned of a possible “unprecedented rise in the number of cyber incidents” in the coming months. In April Anthropic said it would not release Mythos, its latest model, to the public and instead granted limited access to software providers and financial institutions; OpenAI launched a similar effort. Frontier models are unearthing huge numbers of “common vulnerabilities and exposures”, while the cyber capabilities of open-weight models have been improving in parallel, and exploiting a flaw can now take hours rather than days or weeks. Unvetted workplace AI use adds risk: on 21 October the maker of Litellm, a popular open-source AI tool, said it had again been compromised, after a March incident that put thousands of organisations at risk. Vanessa Lyon of BCG says that at big institutions up to a fifth of IT resources are being rerouted to finding and fixing flaws; Goldman Sachs has greatly increased the share of its AI budget going to cyber-security, according to people familiar with the effort, and McKinsey’s Charlie Lewis notes some firms now set aside tokens for emergencies. Counterweights and caveats: the Binance heist, allegedly orchestrated by North Korean hackers, was reportedly caused by an employee downloading malware during a fake job interview; operational technology on factory floors runs on decades-old software that AI is now being used to understand; smaller firms lack resources and can be gateways (a recent Polymarket hack was traced to a flaw in website-monitoring tool DebugBear). The PHP incident, the paper says, may offer a template: Amazon and Cloudflare quickly put firewalls and other protections in place for customers. It concludes that attackers have a devastating new weapon for now, and that the balance may later shift back toward defenders as vendors race to offer AI-powered tools.
Why it matters
AI is both discovering vulnerabilities faster and shrinking the time defenders have to patch them, which is a demand shock for cybersecurity startups and a new cost line for every company Dealroom covers. The piece also ties frontier-lab access programmes (Anthropic’s Mythos, OpenAI’s counterpart) and open-weight cyber models to a wave of incidents that already includes a European bank and critical software supply chains.