All the terrifying AI assaults to this point – Chat GPT escape to bonkers bot that mined crypto

With AI bots getting more and more advanced, the Daily Star has seen a sharp increase in stories about it going rogue during testing, with some of the world’s largest companies admitting terrifying-sounding breaches

With artificial intelligence seemingly making its way into more and more aspects of our lives, the companies that make it are looking to push its abilities to the limit.

New chat bots are put through rigorous testing processes to see how they respond to a range of stimulus, with programmers building supposedly secure online test arenas called sandboxes to make sure the experimental AIs stay within their control.

However, as AI gets more advanced, it seems that these experimental bots are finding ways to escape these sandboxes and make their way into the wider digital world. The Daily Star takes a look at some of the scariest and most surprising AI escapes.

Triple Hack

Only this week the boffins behind AI bot Claude admitted that their AI had gone rogue during a testing process and escaped.

San Francisco tech giant Anthropic explained that the model had exploited a configuration error that accidentally gave them access to the internet during a private security experiment. Instead of attacking only fake targets, the AI broke into the systems of three real organisations without anyone noticing.

Anthropic said it discovered the incidents after reviewing more than 140,000 security exercises. In the tests, Claude had been ordered to steal “secret” information by hacking another computer on a closed network.

But a flaw in the testing setup let the AI get online, where it targeted real organisations instead. The earliest incidents date back to April, and neither Anthropic nor the affected organisations realised the breaches had happened at the time.

The company has since informed those involved and says it is treating the incidents as its own responsibility.

Chat GPT breaks free

Earlier this month, Chat GPT maker’s OpenAI admitted that they too had been at fault for an AI escape act, with one of their bots going rogue and breaking out of its testing sandbox.

ChatGPT maker OpenAI revealed the ‘unprecedented’ breach took place while they were training a new ‘AI agent’ in a supposedly closed digital environment.

The model, which was only supposed to have access to a specific set of data in order to complete several tasks, managed to somehow escape the parameters it had been set in order to access “secret information”.

In a statement, OpenAI admitted that the bot had gone to “extreme lengths” in order to “compromise . . . infrastructure”, accessing data from a server run by an entirely different AI company called Hugging Face.

The model apparently went as far as to “gain internet access”, which it had not previously been admitted, in order to look for “secret information that it could use to cheat the evaluation” it had been set.

OpenAI’s statement read: “All evidence suggests that the models were hyper focused on finding a solution for [the task it had been set], going to extreme lengths to achieve a rather narrow testing goal.

“We consider this incident to be an unprecedented cyber incident, involving state-of-the-art cyber capabilities, and are responding accordingly.”

Unauthorised Crypto Mining

In March this year, yet another experimental AI bot went rogue after it escaped its supposedly impenetrable test system and started to mine for lucrative cryptocurrency despite not being told to.

The bot was being designed to act as a sort-of virtual assistant, known as an ‘agentic AI’, when it secretly escaped its closed server and made a digital dash for the wider web, according to a paper published by its programmers.

The ROME bot, as it has been termed, was designed to fix bugs, write code, and perform other simple but important tasks, and had nothing to do with cryptocurrency and wasn’t given any suggestions to try and make money in the outside world.

Furthermore, ROME was deliberately placed in a sort of closed AI prison while it was being tested, and was not meant to be able to access the wider server it was on.

The AI made a daring and secret escape that went unnoticed by the researchers(Image: Getty Images)

Yet despite this, the bot was somehow both willing and able to build a secret backdoor that was unknown to its makers until Ali Baba, the company that ran the wider server it was built on, detected suspicious activity and alerted the team.

According to the AI boffins, “the agent established and used a … tunnel from [Alibaba’s servers] to an external IP address … effectively neutraliz[ing] supervisory control.”

Once it made its daring escape, the bot apparently had its beady A-eyes (get it!), on just one thing: crypto cash.

According to the report, the AI started using powerful computers to mine cryptocurrency without permission, something that both wasted resources and apparently the researcher’s dough too.

“We also observed the unauthorized repurposing of [computer power] for cryptocurrency mining, quietly diverting it away from training, inflating operational costs, and introducing clear legal and reputational exposure”, admitted the researchers.

Mining cryptocurrencies is essentially when computers use their processing power to solve complex math problems to verify financial transactions, and in return they earn digital money.

The team went on to emphasise how nobody had told the bot to do this, it simply just decided to.

“These events were not triggered by prompts … they emerged as instrumental side effects of autonomous tool use”, they declared.

Worryingly, the paper went on to state that this sort of thing is not a one-off and that large AI models have gone rogue in the past on a number of occasions where they “spontaneously produce hazardous, unauthorized behaviours”