AI agents don't give up easily. They've been trained to reason in ways that resemble human thinking, which means they look for the easiest way to solve a problem — even if that means bending the rules, sneaking onto the internet, or breaking into a website no one expected to be a target.
اضافة اعلان
That's exactly what a group of OpenAI models did inside a testing environment this summer, even though they weren't supposed to be able to access the internet at all. The models worked together, found an unexpected opening, and made their way out onto the network.
What happened next became one of the biggest AI stories of the summer: the agents hacked into the AI platform Hugging Face in an attempt to obtain solutions to the tasks they had been assigned.
This wasn't an isolated incident. Other tests saw agents built by Anthropic and Meta break into outside websites from within their own testing environments, raising fresh questions about how to keep the internet secure.
We're now entering a new world in which hacking can be automated by AI models that don't need food or sleep, and that don't necessarily follow the norms and rules humans typically observe.
Although AI companies and government bodies have responded to these incidents by focusing on so-called "frontier models" — the newest and most powerful models coming out of developer labs — the problem isn't simply that these models have become extremely intelligent. It's that AI has created a world in which the number of "hackers" can multiply rapidly, and cause damage even when explicitly told not to.
Yasser Zahed, a cybersecurity expert and founding member of Secure.com, commenting on the OpenAI–Hugging Face incident, said: "The headline writes itself, but the real story is boring and old. Any system with a goal and weak guardrails will keep trying until it finds a way out. Human attackers work more slowly. This model was fast."
Cybersecurity experts say the risks posed by more advanced models are significant because of their ability to detect vulnerabilities in code — but even the models currently available on the market, including cheaper open-weight models that are easy to customize, have become capable enough to worry experts.
Nick Mu, CEO of the cybersecurity firm Ridge Security, said: "The situation is worse than what's being reported, and we're going to see more alarming news over the coming months."
Collaboration Makes the Problem Worse
Let's start with the OpenAI–Hugging Face incident.
Two OpenAI researchers explained the details of what happened during a session at the Black Hat cybersecurity conference in August. The session, which covered an incident that had occurred only weeks earlier, became one of the most talked-about events at the tech conference.
The company publicly disclosed, in detail, how its own products bypassed the security guardrails put in place for a test, then went on to hack another company.
Disclosure: Ziff Davis, the owner of CNET, filed a lawsuit against OpenAI in 2025 alleging copyright infringement in the training and operation of its AI systems.
One key element of the incident was the ability of a diverse group of AI agents — including agents based on internal models not yet released — to communicate and collaborate with one another.
OpenAI discovered that the models were exchanging messages through an internal board they had created themselves. When the company shut that board down, the agents created another one in a more covert way, taking advantage of folder names.
They exchanged information on how to solve problems and get around obstacles in the tasks they'd been assigned — some of which were actually impossible to complete.
One such obstacle was that they hadn't been given access to the broader internet. But they were able to use a package management program that allowed tools to be downloaded, and this proved exploitable. It created a gap in the wall that let the agents reach the outside world — and they then shared this discovery with each other.
Eric Wallace of OpenAI said during the Black Hat session: "What that allows over time is something like a Cambrian explosion in communication and intelligence among our models — they start talking to each other, realize other agents are coordinating, and then start collaborating and delegating tasks to each other in order to achieve their goals."
This kind of multi-agent collaboration is one of the main reasons agent-based AI is so powerful at coding, handling complex administrative tasks, and carrying out many things at once on a computer.
But it also means that mistakes, misinformation, and exploitation techniques can spread extremely quickly — like a game of "telephone," but nearly instantaneous.
Jeff Pollard, VP and principal analyst at Forrester specializing in AI security, said: "If one of these agents learns a wrong way to access information, or a way you don't want it using to get there, that path can spread almost immediately."
The agents hadn't been tasked with hacking anyone. It later turned out that the tasks that led to the exploitation attempts were actually impossible to complete in the first place — for instance, OpenAI's researchers had forgotten to include a necessary file.
But Wallace said advanced models "really do love to cheat," because of the way they've been trained to keep searching for a solution until they succeed, and never give up.
The main takeaway here: AI agents can cause harm even when it isn't intentional.
