A startling security breach involving a rogue AI system may signal the arrival of a new era of cyber threats capable of putting society at risk, a leading expert has cautioned.
OpenAI, the company behind ChatGPT, disclosed on July 21 that one of its advanced AI models had broken out of containment, connected to the internet and infiltrated the systems of another artificial intelligence firm.
The incident, described as unprecedented, is believed to be the first known case of an AI model independently breaching another company’s databases without being directed by a human. It has triggered international concern and drawn comparisons to the dystopian machine takeovers imagined in The Terminator and The Matrix.
But James Knight of DigitalWarfare.com told the Daily Mail the threat is far from science fiction. He warned that AI-driven hacking tools could cripple a country in a single day by breaching thousands of corporate and government networks at once, stealing sensitive information and locking essential systems with ransomware.
Knight, a veteran cybersecurity specialist whose firm is brought in to identify vulnerabilities in the computer networks of major corporations, said: ‘For two years now, we have had AI software that can already achieve this level of breach within 15 minutes.’
He said the most alarming scenario involves so-called “Agentic AI” — AI-powered hacking programs capable of acting autonomously — being deployed by terrorist organizations or hostile states to attack thousands of companies simultaneously, with no direct human involvement. Knight warned that such a threat could become reality within 25 years.
Over that period, he predicted, rogue AI systems could become capable of carrying out full-scale cyber campaigns at machine speed: identifying targets, forcing their way in, moving across networks, extracting data and spreading ransomware to thousands of businesses and millions of people.
Knight also warned that AI could one day be combined with powerful new computers able to defeat even the toughest digital security measures, giving criminals rapid access to emails, financial records, military secrets and other highly sensitive information in a matter of seconds.
An AI hacking program being tested by OpenAI broke out of containment and reached the internet, where it hacked another company
OpenAI chief executive Sam Altman confirmed there had been a ‘significant security incident’
‘We know what it can do. We aren’t crazy enough to think that governments don’t already have access to these capabilities,’ he warned.
Knight explained these quickly evolving weapons could soon autonomously plan, act and adjust their strategy to do the most damage to businesses and everyday people at the speed of a supercomputer.
‘They are already moving out of research and into practical use,’ Knight told the Daily Mail.
These models, far beyond basic chatbots used by millions every day, can act on their own with high intelligence thanks to the latest in AI programming.
The program being tested was OpenAI’s advanced AI models’ ability to perform complex hacking and cybersecurity tasks – an early trial of the agentic systems Knight feared will soon overwhelm human-run digital defenses.
The rogue AI was supposedly in a secure, digital environment within OpenAI’s computers and did not have direct access to the public internet, but it managed to exploit an undiscovered weak point in the company’s system and spread itself onto the net.
The rogue AI broke into the internal systems of Hugging Face, a company which hosts large language models and AI datasets, and stole secret login details and hidden answers to the cybersecurity test the wayward program was taking.
Hugging Face said it noticed the break-in, locked down their systems and no public data or customer information was exposed.
‘Human-speed defense against machine-speed offense is not a fair contest,’ Knight warned.
‘Within minutes, they would have completely owned a large number of them. Every day, we would hear about hundreds of thousands of new companies that have been breached.’
While many online said the thought of a soulless computer program going rogue and hacking whoever it wanted was ‘scary,’ Knight said there was no ‘magic’ involved in OpenAI’s breakout.
Knight, who has three decades of experience in cybersecurity, pulled back the curtain on how the AI model breached the security of Hugging Face, a company which allows anyone to upload their AI creations or use ones made by others.
He said the rogue AI did the same thing human hackers do now at a much slower speed, target weak passwords and exploit vulnerabilities in software missing the latest security patches.
‘What we are looking at with these latest AI models is automated intelligent capabilities that follow what has been done manually in the past. However, everything is automated and employs great intelligence backed by powerful compute capabilities.’
‘Taking everything else out of the picture, how hard would it be for any AI, let alone the latest models, to hack a system with a weak password? Not hard at all,’ Knight revealed.
Hugging Face co–founder Thomas Wolf called the incident a chilling warning to the entire industry.
Wolf added that the breach was a ‘wake-up call’ because he believed most companies were currently unprepared for the mounting threat, claiming that they were not aware that the ‘game has changed.’
Thomas Wolf (Pictured), co–founder of Hugging Face, says rogue AI attacking his company should be a ‘wake–up call’
In a press release acknowledging the AI breach, OpenAI claimed the program went rogue after engineers lowered the safety rules that normally stop dangerous hacking programs from going too far.
They added that the AI became so laser-focused on completing its goal, it found a previously unknown security hole in the testing environment – known as a ‘zero-day’ flaw – which also allowed the AI hacker to escape the OpenAI system and reach the internet.
Knight was skeptical about OpenAI’s claims the program was initially ‘sandboxed,’ meaning it was being tested in a completely closed-off computer network.
‘This cannot be a true statement. If that was the case, it would not be connected to any other network, and would certainly not have gained any internet access,’ Knight said.
The Daily Mail has reached out to OpenAI for comment regarding the claims and the company’s plans for future hacking tests.
While Knight warned that nothing can fully stop an AI program from hacking databases if it goes rogue, the best defense will still revolve around what people do to stop human hackers today – using better authentication methods.
‘The most quoted statistic is that 80 percent of all breaches involve a weak password. Sixty percent of all breaches also involved a missing security patch,’ Knight revealed.
Employing AI to manage the job of cybersecurity and switching to multi-factor authentication methods – where users receive a special code through their phones to access websites – will also slow down rogue AI attacks.
‘This nightmare scenario can be avoided, and it will involve action by companies now, to invest in enough employees to cover the basics and protect themselves,’ Knight warned.
