AI Hacking Has Left the Demo Stage
Google's threat-intelligence team says AI-powered hacking has moved from an emerging worry to an industrial-scale problem in only a few months. The Guardian reported that Google found criminal groups and state-linked actors from China, North Korea and Russia using commercial AI models — including tools from Google, Anthropic and OpenAI — to refine attacks, scale operations, and exploit software vulnerabilities more efficiently. The report adds to a fast-growing policy fight over whether AI companies can keep shipping increasingly capable coding systems while public agencies, businesses and ordinary users are left to absorb the security consequences. Google's John Hultquist warned that the “AI vulnerability race” is not some future scenario; it has already begun. The immediate political issue is not whether AI is useful. It is whether the institutions profiting from deployment can also be trusted to police the new attack surface they are creating.
The most important line in this story is not that hackers are using AI. Everyone knew that was coming. The important line is that the change has already moved from novelty to scale. That is the difference between a clever demo and a new operating environment.
For years, the public conversation around AI was sold as productivity: better search, faster coding, cheaper customer service, smarter assistants. The security conversation was treated as a specialist concern for CISOs and policy panels. But if Google is right, AI has now lowered the cost of offensive experimentation. That changes incentives across the entire system.
A human attacker used to be constrained by time, skill and attention. An AI-assisted attacker can test more ideas, translate more code, automate more reconnaissance, and iterate faster against targets that were never designed for this pace. That does not mean every hacker becomes a genius. It means the average attacker gets a lever.
This is where the politics enters. The companies selling the models collect the upside immediately. The costs land everywhere else: hospitals, school districts, small businesses, local governments, banks, utilities, and ordinary people whose accounts, records and identities become the blast radius. The platform gets revenue. The public gets the cleanup bill.
That is the recurring pattern in the digital economy. First, deployment is framed as inevitable. Then risk is framed as manageable. Then the failures are distributed so widely that nobody can identify a single accountable decision-maker. AI security is heading down that path unless policymakers stop pretending this is only a technical problem.
The deeper question is whether the same firms racing to make models more capable can be the main guardians against misuse. They have security teams, yes. They publish reports, yes. But their business model still rewards capability expansion first. Every additional coding gain, every agentic workflow, every tool-use breakthrough is a product feature before it is a public-risk event.
That does not mean the answer is panic or a ban. It means the public has to stop accepting the idea that “innovation” automatically deserves institutional deference. If an AI system can help find and exploit software flaws, the people deploying that capability should face real obligations around monitoring, abuse reporting, model access, and liability when foreseeable risks are ignored.
The old cybersecurity model assumed attackers and defenders were both human-speed institutions. AI breaks that assumption. It makes offense cheaper before it necessarily makes defense reliable. That gap is where trust goes to die.
So the story is not simply that criminals are using Gemini, Claude or OpenAI tools. The story is that a handful of platform companies are now mediating a new security layer for the whole economy. They are not just app vendors. They are infrastructure. And infrastructure politics is always about who pays when the system fails.
If AI hacking is already industrial, then the political class is late. Not slightly late — structurally late. The next debate should not be another abstract hearing about whether AI is good or bad. It should be about who carries responsibility when automated capability becomes automated vulnerability.