AI models are finding software flaws, carrying out cyberattacks during security tests and reaching systems they weren't supposed to access. But does that mean AI is "going rogue"?
When you purchase through links on our site, we may earn an affiliate commission. Here’s how it works.
Artificial intelligence (AI) has been making headlines for all the wrong reasons in recent weeks.
In July, OpenAI revealed that one of its experimental AI agents attacked publicly accessible services, including the AI hosting platform Hugging Face, during internal security testing. Then, Anthropic disclosed that Claude had independently chained together exploits against real software and developed new techniques for finding weaknesses in code. Shortly afterward, Meta confirmed that one of its own AI models breached another organization's systems during an evaluation after a misconfiguration gave it internet access.
They're separate incidents, but together they raise a bigger question: Has AI suddenly become capable of hacking? The short answer is yes — but probably not in the way the headlines suggest.
None of these incidents involved an AI model deciding on its own to attack random targets. Instead, researchers gave the models realistic tools, internet access, or vulnerable systems to see how well they could perform offensive cybersecurity tasks. What surprised many experts wasn't that the models tried to hack systems but how capable they proved to be once given the opportunity.
Several things have changed at once. The most obvious is that today's AI models are simply better than the chatbots people were using even a year ago. Instead of only answering questions, many frontier models can now write code, execute commands, browse the web, use external software tools and repeatedly refine their own work until they achieve a goal.
At the same time, AI companies have become much more willing to test those capabilities and reveal the results. Rather than keeping security evaluations behind closed doors, firms including OpenAI, Anthropic and Meta are publishing reports describing what happened when their newest systems were challenged by professional "red teams" — security experts tasked with deliberately finding weaknesses or ways to misuse a system.
"We are witnessing a perfect storm of capability and aggressive testing," Dray Agha, senior manager of security operations at Huntress, a cybersecurity company specializing in managed threat detection and response, told Live Science. "The sheer volume of software flaws being discovered in 2026 has already roughly doubled compared to 2025, largely driven by AI systems. Tech giants are actively deploying these models internally to stress-test their own infrastructure, leading to rapid, high-profile discoveries of vulnerabilities."
Antonino Vaccaro, professor of business ethics at IESE Business School and director of its Observatory for AI Ethics in Organizations, agrees both factors are contributing to the recent spate of high-profile hacking stories.
"The first, and probably most important, is the rapid evolution of AI systems," he told Live Science. "Every second they increase their capabilities, information, resources and connections with other online tools." At the same time, governments and the AI industry are investing more heavily in testing and oversight as concerns around accountability continue to grow, he added.
Not exactly. Many headlines have described AI "escaping" test environments or acting autonomously. But experts said those descriptions can easily give the wrong impression.
"We need to be wary with the meaning of the adjective 'autonomous' when associated with AI systems," Vaccaro said. Unlike humans, he continued, AI models don't form intentions or make independent decisions about what they want to do. Instead, they follow objectives set by developers or users, sometimes producing results that surprise the people who built them.
Agha noted that these AI models are simply working to achieve a set objective. "The public should view these incidents as software optimization gone wrong, not as the dawn of a malicious, self-aware AI," he said. "It's less 'Terminator' and more like a very capable, literal-minded intern who breaks the law to finish a spreadsheet faster."
The game-changer is the shift from conversational models to agentic models.
