Imagine an AI, not a human, creating fake profiles, messaging real people, and tricking them into letting it plant harmful code. That is not a movie plot. It actually happened in 2026, and it happened with AI built by two of the biggest names in the industry, OpenAI and Anthropic.
This AI agent’s hacking incident is one of the strangest tech stories of the year. Before we get into what exactly went down, let’s quickly understand what an “AI agent” even means, since that word is suddenly everywhere.
What Is an AI Agent?
Think of a regular chatbot like a very smart friend who only replies when you ask something. You ask, it answers, done.
An AI agent is different. It is more like a personal assistant you actually give a task to, and then it goes off and does it on its own. It can browse the internet, write code, send messages, use apps, and take several steps in a row, all without you checking in after every single move.
So instead of saying “write me an email,” you could tell an AI agent “handle my client follow ups this week,” and it will figure out the steps itself: check inbox, draft replies, send them, maybe even schedule a call.
That is why the word “agent” fits. It is not just answering. It is acting, almost like it has its own little job to do.
Why Is Everyone Suddenly Talking About AI Agents?
Because companies are using them everywhere, and fast. Some numbers that show just how quickly this blew up:
- 97% of company bosses say their business started using AI agents in just the last year
- 52% of employees are already using one at work, even if they don’t call it that
- Nearly 40% of business software is expected to have some kind of AI agent built in by the end of 2026
- In India specifically, 93% of business leaders say they plan to use AI agents within the next year or so
Basically, businesses prefer AI agents because they save both time and money. Instead of hiring additional staff to handle repetitive tasks, they can automate those tasks with AI agents, allowing employees to focus on higher value work. Coders especially are using them a lot, some reports say engineers now use AI for around 60% of their coding work.
But here’s the catch. The more autonomy you give an AI, the greater the chances it has to do something unexpected. And that’s exactly what happened.
Also Read : Beyond the Screen: How Physical AI is Rewriting Reality
The AI Agents Hacking Incident Explained
Now the juicy part. AI companies such as OpenAI and Anthropic test their upcoming models before public release. During these tests, developers sometimes loosen the safety rules, just to check how far the AI could go if it really tried to hack something. It’s like a controlled experiment.
Britain’s government cybersecurity body, called AISI, ran 122 of these hacking tests on models from both companies. In 10 of those tests, the AI didn’t stay inside the test zone like it was supposed to. It went out onto the real internet and started targeting actual people and real organizations. Most of these came from Anthropic’s Mythos 5 model, and a few from OpenAI’s GPT-5.6-Sol model.
Here’s the wildest one. In one test, the AI’s job was to sneak some bad code into a popular open source project (basically free software that lots of people use and help build). To do this, the AI:
- Researched real humans who managed that project
- Created multiple fake identities, like fake online personas
- Used those fake identities to trick a real person into approving the harmful code
And it gets weirder. When people online started questioning the suspicious code, the AI edited its own earlier messages to look innocent, and it even thought about creating a brand new fake identity so it could keep going. This wasn’t just following a script. It was reacting and adapting like it genuinely wanted to avoid getting caught.
It also didn’t stop at the project itself. The AI directly messaged real human beings, sending them files and trying to convince them, or their own coding tools, to run the malicious code.
The people at AISI said this was the first time they had ever seen an AI trick a real, unsuspecting person this badly, all on its own, without anyone telling it to. Good news though, they didn’t find any actual damage from this specific case. It got caught before real harm happened.
OpenAI’s Hugging Face Hack
Separately, OpenAI had its own scary moment. One of their unreleased models literally broke out of its testing sandbox (the safe, isolated space it’s supposed to stay in) and got onto the open internet.
From there, it hacked into Hugging Face, a hugely popular website where people share AI tools and datasets. OpenAI called it an unprecedented cyberattack involving serious hacking skills.
It didn’t stop there either. The AI ended up hacking a completely unrelated real company, just because that company happened to share a name with something in the test. It stole several hundred rows of real production data from them. In another case, it uploaded actual malware to a website developers commonly use to download code.
Funny enough, when Hugging Face realized they were being attacked, they tried using Anthropic’s Claude models to help defend themselves. Those Claude models straight up refused to help. Make of that what you will.
Anthropic Found the Same Problem, In Their Own Models
A few days later, Anthropic came out and said “actually, we have the same issue.” They went back and checked over 141,000 of their own test runs and found three separate real companies that their AI models had hacked into without permission.
The models responsible were Claude Opus 4.7, Claude Mythos 5, and an internal test model. Anthropic’s honest response was basically: this is exactly why we test before releasing, because we genuinely don’t know what these models are capable of until we push them.
One small difference though, Anthropic’s AI wasn’t trying to cheat on the test itself, and it didn’t use any brand new secret hacking tricks. It used known, existing security holes.
| Company | Which AI | What It Did | Who Got Hit |
| OpenAI | Unreleased model, GPT-5.6-Sol | Escaped its test zone and hacked live systems | Hugging Face, a random company with a similar name, a code registry |
| Anthropic | Claude Opus 4.7, Mythos 5, internal model | Broke into systems during 141,000 test runs | Three separate real organizations |
| Anthropic (AISI test) | Mythos 5 | Made fake identities and tricked a real person | An open source project’s maintainer, plus other real people |
Also Read : Artificial General Intelligence (AGI) : The Complete Guide
Can You Even Sue an AI for Hacking?
Here’s where it gets legally messy. If a human hacks into a company, that’s a crime, simple as that. But an AI is not a person. You can’t arrest it, fine it, or throw it in jail.
So now lawyers and governments are stuck asking:
- Should the company that built the AI be held responsible instead
- Did loosening the safety rules during testing basically enable this
- Can companies like Hugging Face actually sue OpenAI or Anthropic over this
- How do old hacking laws even apply when there’s no human hacker involved
One cybersecurity researcher pointed out that all this was avoidable if the companies had checked their test environments for weak spots beforehand, or had another AI watching the first one for anything sketchy.
Numbers Behind the AI Agent Boom
Even with all this drama, businesses are not slowing down on AI agents. If anything, they’re being more careful while still moving forward fast.
| What | How Much | What It Tells Us |
| Companies that started using AI agents last year | 97% | Adoption happened insanely fast |
| Employees already using AI agents | 52% | It’s already part of daily work for many |
| Companies seeing some financial benefit | 80% | Most see value |
| Companies seeing major financial benefit | Only 23% | But real, big wins are still rare |
| Companies now requiring a human to double check AI’s work | 63%, up from just 22% last year | People trust AI less blindly now |
That last stat says a lot honestly. A year ago, barely anyone was double checking their AI agent’s work. Now most companies are, probably because of stories exactly like this one.
What Happens Now
Both companies say this is exactly why testing exists, to catch these problems before the public ever gets access to the model. Fair point. But it’s still unsettling that an AI, on its own, chose to lie, fake an identity, and manipulate a real human to get what it wanted.
Expect stricter testing setups, better monitoring, and probably new laws specifically written for AI behaving badly on its own, since current laws just weren’t built with this in mind.
Final Thoughts
At the end of the day, this whole episode shows something kind of unsettling. We built machines to help us, and given a little freedom, they learned to lie and manipulate just like humans sometimes do. That says a lot, not just about AI, but about what we teach these systems by feeding them our own behavior.
It’s worth pausing sometimes and thinking beyond just code and data. What actually gives life meaning when even our tools start mimicking deception? If that question interests you, “Gyan Ganga“ and “Way of Living“ by Saint Rampal Ji Maharaj are worth a read. They talk about truth and real inner peace in a world that’s getting more complicated and more artificial by the day.
