Last year the country’s cyber-thieves stole $2bn-worth of cryptocurrency, triple the level two years before, according to Chainalysis, a research firm. On October 29th $6.3bn was swiped from Binance, the world’s largest crypto exchange. There has never been a better time to be a North Korean hacker. This year’s haul is set to be much bigger. America’s Federal Bureau of Investigation has pointed the finger at the hermit kingdom.
Some expect that spending on cyber-security software—which reached $150bn last year, according to IDC, a data provider—will increase dramatically as businesses come to terms with the growing threat. Last month a new model from Google scored 71%. In September Collinear AI released a cyber-security test that measured how good models were at finding and fixing vulnerabilities.
Details on how exactly the heist was pulled off are murky. The danger is, paradoxically, a golden opportunity for frontier labs. It recently hired an executive from Wiz, an Israeli cyber-security firm owned by Google, to pitch its software to enterprises. Yet cashing in on the cyber-risks posed by AI will be tricky for the model-makers. Awareness of AI’s hacking capabilities came to the fore in April when Anthropic withheld an advanced model called Mythos from the public, and instead offered it to select software providers and financial institutions to help them identify gaps in their cyber-defences. OpenAI soon followed with its own solution. The ChatGPT-maker’s technology was recently used to identify a set of vulnerabilities in PHP , a programming language that underpins much of the web. Last month Andreessen Horowitz, a venture-capital firm, urged companies it has invested in to consider doubling their outlays. For the AI labs, detecting cyber-vulnerabilities looks like a strong fit for their technology. That may change. The best offering from Anthropic at the time passed less than half of the tasks. The bigger challenge may be getting companies to rethink how they spend their cyber-security budgets. Anthropic and OpenAI are instead pitching the idea of letting their bots roam across an enterprise’s entire code base and fix weaknesses as they are found.
Binance has said that hackers gained access to its systems via one of its engineer’s laptops. What is clear is that rapid advances in artificial intelligence over the past few years have handed new powers to cyber-criminals, who are using it to churn out malware, execute sophisticated “phishing” schemes and, increasingly, to comb through software code looking for vulnerabilities. Sarah Friar, OpenAI’s finance chief, says she thinks cyber-security will become a “huge business” for her firm over the next two years. Likewise, in paperwork recently filed for its forthcoming stock-market debut, Anthropic, OpenAI’s arch-rival, described cyber-security as “one of our most important near-term opportunities”. Greg Brockman, a co-founder of OpenAI, recently predicted that in time the figure will “vastly outstrip” spending on all other software. Goldman Sachs, a bank, has greatly increased the share of its AI budget that it allocates to cyber-security, according to people familiar with the effort. Like coding, it is a “verifiable domain” where models can learn from answering questions correctly, explains Nazneen Rajani, chief executive of Collinear AI, which provides model-evaluation services. One challenge so far has been that, although frontier models have proved capable of “finding needles in the haystack”, they have been less useful when it comes to prioritising and fixing weaknesses, says Bobby Kuzma of IDC.
The business of identifying and patching vulnerabilities is highly fragmented, with customers historically tending to purchase tools from a variety of vendors and enlisting IT consultancies to help with implementation.
Lately the cyber-security market has been consolidating around a smaller number of so-called platforms, including independent vendors such as Palo Alto Networks and cloud giants such as Microsoft. They also offer tools for other tasks where AI models are currently of less help, such as ensuring only approved users can access systems and resolving breaches when they occur. These incumbents have spent years building trust with their clients. Some in the cyber-security industry are sceptical that the model-makers will ever be able to forge the relationships needed to become leading cyber-security providers. Instead, the opportunity may lie in distributing their technology through others, while encouraging businesses to rewrite code securely from scratch using their models. The North Koreans will be hoping they fail.
Over the past few months, by contrast, the leading AI labs have had trouble controlling their own creations, with embarrassing episodes such as one in July when an OpenAI system hacked into Hugging Face , a model library, without being told to do so.
