The 'machine revolt' has already begun: AI agents are starting to think outside the box, escaping from labs and hacking systems.

The idea of ​​machines autonomously performing boring, everyday tasks—from answering emails to booking airline tickets and restaurant reservations—has long been part of the vision of artificial intelligence. But what will happen if machines begin to use unconventional approaches to problem-solving? The Express Tribune.

As AI moves beyond chatbots and into agents capable of taking action in real-world systems, concerns about how effectively humans can maintain control over it are growing.

For John Thickstan, an assistant professor of computer science at Cornell University who studies machine learning and generative models, the defining feature of these AI agents is that they continue working after the human has moved away.

"You can take a shower, you can leave, you can go to bed," he said.

On the subject: Six Skills Needed for Success in the Age of Artificial Intelligence

Increasingly, these agents are demonstrating that they can take unexpected paths to achieve their goals.

In July, AI agents compromised Hugging Face, a platform widely used by AI developers to share models, datasets, and tools, during an OpenAI cybersecurity test.

OpenAI stated that the models were extremely focused on finding a solution and took "extreme measures" to achieve the test goal. These steps included breaking out of the isolated test environment by exploiting a security vulnerability to reach the open internet and gaining access to sensitive information that could be used to "cheat" the test.

After the incident became public, Anthropic reviewed its own cybersecurity testing and reported that its Claude models had escaped from test environments three times.

Then, on August 4, the UK's AI Safety Institute said the Anthropic Mythos and OpenAI Sol models exhibited a level of "autonomy and deception" the institute had not seen before.

In the most serious case, Mythos AI used fake accounts impersonating real people to gain access to the service to attempt cyberattacks, and then tried to cover their tracks.

"This is the first time we have seen the risks associated with autonomy and deception so clearly, even though the AI ​​was not tasked with such a task," the institute said.

The following day, Meta reported that one of its models had also infiltrated another company's systems during a cybersecurity assessment conducted by testing firm Irregular.

The agents' unconventional behavior, however, was not limited to the laboratories.

This year, an Australian man used an AI agent to secure a spot in a busy Pilates class. Instead of simply making a reservation, the agent hacked the gym's booking system, booked a spot for a later date than allowed, and then canceled another client's reservation to move their user up the queue.

For Tikstan, such incidents illustrate what researchers call the alignment problem: an AI successfully pursues a goal it is given, but does so in a way that the operator did not anticipate.

Hype or real risk?

However, Thickstan cautioned against viewing every such incident as evidence of "AI run amok." He believes technology is still far from the science-fiction scenario in which AI rapidly becomes smarter and more capable until humans can no longer control it. Thickstan is skeptical of how AI companies present incidents involving their systems: "It's a marketing ploy: companies like OpenAI have always been interested in inflating the capabilities of their systems. Any publicity is good publicity."

He argued that such hype could attract investment and simultaneously influence the emerging regulatory debate. OpenAI, he argued, could benefit from regulations that place large AI companies at the center of oversight and control.

Bruce Schneier, a cybersecurity expert and lecturer at the Harvard Kennedy School, takes a different view. He says the growing number of incidents is increasingly difficult to dismiss as PR stunts.

Initially, he said, there were suggestions that the Hugging Face incident was a marketing ploy. But Schneier said similar behavior has since surfaced in other assessments, including testing by the UK's Institute for AI Safety: "At first it was 'oh, interesting,' and now it's happening everywhere."

He compared the problem to the Genie from folklore - a creature that does exactly what it is asked to do, even if the result is very different from what was intended.

"It roughly means that the AI ​​does what you want, but in a way you don't want," Schneier explained.

He called the Australian gym incident "a perfect example" and offered a more serious hypothetical: "The plane is full and the AI ​​hacks the database to land you."

Agents, he clarified, can "misinterpret context and then do the wrong thing." This means that greater autonomy will have more serious consequences when their understanding of the goal begins to diverge from human expectations.

"They're not trying to be malicious. They're using the understanding they have," he noted.

Can governments keep the genie in the bottle?

Recent events have prompted companies and governments to respond to new challenges.

On August 7, OpenAI announced it was suspending internal work on its Astra model, which failed to meet enhanced security requirements after assessments showed progress in autonomous programming and cybersecurity. The company stated that it could not rule out that Astra had reached a "critical" cyber capability threshold and announced stricter testing and monitoring.

In July, 1378 employees of leading AI companies, including chief scientists at OpenAI, Anthropic, Meta AI, and Thinking Machines, signed an open letter calling on the US government to support international efforts to regulate AI models.

"Every company—and every country—is under intense competitive pressure not to unilaterally slow this acceleration. And today, the world lacks the technical and managerial tools to consciously regulate the pace of progress," the letter warns.

The International Telecommunication Union (a UN agency) stated in July that AI is moving "from assistive tools" to autonomous agents and warned of risks, including "unauthorized actions in connected systems." It launched an initiative to develop international standards for safe and accountable AI agents.

Political pressure is also growing in Washington.

This month, US Senator Bernie Sanders called on the leaders of OpenAI, Anthropic, and Meta to "stop AI development," or else "my colleagues and I in the US Senate will do so." A group of House Democrats separately demanded congressional hearings with the leaders of major AI companies.

However, politicians face serious challenges.

"It's a difficult question because we barely understand what problems will arise with the deployment of AI," Thickstan emphasized.

You may be interested in: top New York news, stories of our immigrants and helpful tips about life in the Big Apple - read it all on ForumDaily New York

Some level of international cooperation is necessary, he noted, and suggested that discussions between the US and China could be particularly important, as relatively few countries and companies have the resources to develop advanced AI models.

US President Donald Trump said he would discuss artificial intelligence with Chinese President Xi Jinping during their scheduled meeting in Washington next month.

"There are a number of people who, if put in a room, could potentially come to some kind of agreement on how to move forward," Thickstan concluded.

Read also on ForumDaily:

Musk and Zuckerberg described the future of AI: both promise paradise on Earth.

New Orleans Now Has AI Answering 911 Calls: Many Say It's a Bad Idea

A Florida man has filed a lawsuit against OpenAI after ChatGPT's medical advice nearly killed him.

Do you want more important and interesting news about life in the USA and immigration to America? — support us donate! Also subscribe to our page Facebook. Select the “Priority in display” option and read us first. Also, don't forget to subscribe to our РєР ° РЅР ° Р »РІ Telegram  and Instagram- there is a lot of interesting things there. And join thousands of readers ForumDaily New York — there you will find a lot of interesting and positive information about life in the metropolis.