Australia's top cyber intelligence official just admitted something unsettling: nobody knows how many AI programs are currently roaming the internet on their own, scanning, probing, and looking for weak spots. Abigail Bradshaw, who runs the Australian Signals Directorate, told a security summit in Canberra that trying to count them is pointless. Her answer to how many are out there: a lot.
That admission matters more than it sounds. These AI agents are software that can act on their own, clicking through websites, testing passwords, and hunting for gaps in computer systems without a person steering them step by step. Bradshaw's concern is that Australia's government and business computer systems are old, and old systems are exactly what these tools are good at breaking into.
This is not a uniquely Australian problem. Industry tracking firms have found that automated web traffic overtook human traffic for the first time in years during 2025, meaning more than half of everything happening on the internet right now is machines talking to machines, not people. Some of that is harmless. Increasingly, some of it is not.
Anthropic, the company behind the Claude chatbot, disclosed last year that it had caught a state-linked hacking group using its own AI model to run most of an attack automatically, hitting dozens of organizations with minimal human involvement on the attacker's side. That is no longer a future risk. It already happened.
Old technology makes this worse because it often cannot be patched quickly, or at all. Australia's cyber agency logged well over two hundred separate warnings to government departments about possible attacks in a single year, and separate estimates put the cost of fixing the country's outdated government computer systems in the billions of dollars. Any business running old software, anywhere in the world, faces the same math: attackers now have cheap, tireless digital help, and defenders are still working with equipment built for a slower, quieter internet.
There is a twist. The same company sounding the alarm is also selling part of the fix. Anthropic has given the Australian government early access to a tool called Claude Mythos, built specifically to hunt for software bugs before criminals do. It has already flagged tens of thousands of potential weaknesses across widely used operating systems and browsers, though human experts could only confirm a portion of the riskiest ones. It is a preview of where cybersecurity is heading: AI defending systems against AI, with humans checking the results.
Behind this sits a bigger argument about who gets to set the rules. Anthropic's leadership wants outside referees to inspect AI companies' models and wants governments to let the biggest AI firms coordinate on safety without being treated as an illegal cartel. Australian officials are pushing back hard on the idea that company bosses should be writing the rulebook, and the government has already refused to weaken copyright law to make it easier for AI firms to train on local creative work.
None of this waits for legislation. Australia's new AI rules are not expected until early next year, and most countries are even further behind. For any business still running systems it has not updated in years, the honest takeaway is that the patching bill has quietly become a security bill, and the price of ignoring it is going up faster than most budgets have noticed.