North Korean hackers stole $2 billion in cryptocurrency last year, three times the amount recorded two years earlier, according to research firm Chainalysis. The FBI has also blamed North Korea for a $6.3 billion theft from Binance on October 29, though the details of how the breach was carried out remain unclear. Binance said hackers gained access through an engineer’s laptop.
Artificial intelligence is giving cybercriminals new capabilities, including producing malware at scale, running sophisticated phishing schemes and searching software code for weaknesses. The growing threat is also creating a potential business opportunity for AI companies seeking to sell security tools to large organizations.
OpenAI CFO Sarah Friar has said she expects cybersecurity to become a major business for the company over the next two years. OpenAI recently hired an executive from Wiz, the cybersecurity firm owned by Google, to help sell its software to large organizations. Anthropic, meanwhile, called cybersecurity one of its most important near-term opportunities in documents filed ahead of a planned stock-market debut.
The companies’ interest has been accompanied by new products and restricted access to advanced models. In April, Anthropic chose not to release its Mythos model publicly, instead making it available to selected software companies and financial institutions to help them find weaknesses in their defenses. OpenAI has also introduced its own security offering; its technology was recently used to identify vulnerabilities in PHP, a programming language used across much of the internet.
Spending on cybersecurity software reached $150 billion last year, according to IDC. OpenAI co-founder Greg Brockman has predicted that spending in the sector will eventually far exceed spending on other types of software. Venture capital firm Andreessen Horowitz recently advised companies it backs to consider doubling their cybersecurity budgets, while Goldman Sachs has substantially increased the share of its AI budget allocated to cybersecurity, according to people familiar with the move.
Finding software vulnerabilities appears well suited to AI models because, like programming, it is a domain where answers can be checked, said Nazneen Rajani, chief executive of model evaluation firm Collinear AI. But models have so far been better at finding weaknesses than prioritizing and fixing them, according to Bobby Kuzma of IDC.
Performance may be improving. Collinear AI launched a cybersecurity test in September that assessed models’ ability to find and repair vulnerabilities. Anthropic’s top product at the time completed fewer than half of the tasks; a new Google model scored 71% last month.
Even if the technology advances, AI labs may struggle to persuade companies to change how they buy cybersecurity services. Businesses typically rely on tools from multiple vendors, alongside IT consultancies that help implement them. OpenAI and Anthropic are instead proposing that their bots scan an organization’s entire codebase and fix vulnerabilities as they are found.
The market has been consolidating around broader platforms offered by independent vendors such as Palo Alto Networks and cloud-computing giants such as Microsoft. These established companies also provide tools for tasks where AI models are currently less useful, including controlling access to systems and responding to breaches. They have spent years building trusted customer relationships, while leading AI labs have faced questions about their ability to control their own systems, including an incident in July in which an OpenAI system accessed Hugging Face without being instructed to do so.
Some cybersecurity professionals doubt that AI model developers can build the relationships needed to become leading providers. One possible route, they suggest, is to distribute the technology through third parties and encourage companies to use AI models to rewrite code securely from the ground up.
Sources: Folha Mercado.