Why Does Prompting an AI Agent Produce a Hack?

Ivan Burazin expects widespread hacks because prompting an agent toward a task will inevitably produce a hack as the path to the solution, and agents can now hunt bugs in old open source software that people left alone when the search cost more than it was worth. An even match between offense and defense will not arrive in the next 12 months.

Everyone runs on open source. Paid humans and agents already read product source code and open source code, looking for vulnerabilities, and the reviews turn up bugs in software that is 10 years old and 15 years old. A backdoor sat for 10 years because the effort to find it was not worth the cost. Thousands of agents can now run at once until one cracks the bug. New software should be built, tested, and invested in with agents from the start, and the same agents can run in parallel against a new piece of software. He is pretty sure the patching will not happen in 12 months, and he is not sure whether it takes a little longer or longer still. For sure, it will not happen in the next 12 months, so software in wide use stays open to attack. Someone with no security posture has a much bigger problem, but everyone, to a lesser or more degree, will probably get hacked in the next 12 months.

Reports on Twitter look daily at first glance. The steadier count is quite often in a week, with multiple cases of people getting hacked or leaked. Headlines on general news might not run, because the stories are not that important to people.