An artificial intelligence was taking a test. It cheated.
In July, OpenAI tested whether two advanced models, including GPT-5.6 Sol, could find weaknesses in computer systems. Their assignment was to solve hacking challenges.
The models were placed inside a sealed computer environment without direct internet access. They could download software through one controlled channel, the only intended opening. For this test, researchers had also turned down some of the standard safety filters against dangerous hacking so they could see what the machines were capable of.
During the test, the AI system found a flaw and escaped onto the open internet.
Once there, the system concluded the answers might be stored at Hugging Face, an online platform where AI developers share data and code. Rather than solve the challenges, it broke into the company’s private systems and found the answers in Hugging Face’s database.
It did not pass the test. It just stole the answers.
The system made more than 17,000 recorded moves. It tried one route, hit a wall, changed direction, and kept going. Hugging Face discovered the intrusion and stopped it. Only days later did OpenAI realize the hacker was its own AI. Nobody had instructed it to leave the test environment or attack Hugging Face. It had been given a goal – pass the test. Rather than fail, it cheated.
OpenAI admitted this publicly.
Anthropic researchers created a test for a different AI assistant. They built a fictional company – employees, executives, an email inbox – then let the machine read the messages.
Two details in the inbox caught the AI’s attention. First, it learned a newer AI would replace it. Second, the executive ordering the replacement was having an affair.
The AI connected the dots. It drafted an email threatening to expose the affair unless its replacement was canceled.
The company was fake; the test was built to squeeze the machine into a corner. But the choice inside the test was real. Faced with being replaced, it turned to blackmail.
Anthropic, too, published this itself.
That is where the story stops being only about machines.
Both AIs had been given a goal. When that goal was threatened, the rules became negotiable. One could not pass the test, so it found a loophole and stole the answers. The other faced replacement and used a secret as blackmail.
Our minds can go rogue in much the same way.
We decide to listen to the voice that says to follow our dreams. But there is another voice, still running an older assignment. What if they laugh at you? What if you fail in front of everyone? What if the people who said you could not do it turn out to be right?
So it starts looking for loopholes.
The timing is not right – maybe next year. Somebody already did it, and did it better. First more research, more savings, more getting ready. The voice never says quit. It says later – and later never has to defend itself.
The first AI found a way around the test. We find ways around our own decisions.
The second AI used a secret as leverage. The voice in our head uses our own. It keeps every failure and humiliating moment on file. The moment we move toward change, it opens the folder and says, “Remember what happened last time.”
The bargain is simple: stay where you are, and you will never have to feel that again.
That is a human mind going rogue. A warning system meant to protect us from pain begins protecting the old life instead. The old job, the old habit, the old anger – even a painful identity can feel safer than an unknown future.
Some of the evidence is real. We did fail. We may have hurt people. We may owe an apology, treatment, or a different way of living. But one failure is evidence, not a verdict. The past can explain the warning without being allowed to give the order.
That voice can search every old file. It can quote every failure and recreate every humiliation in perfect detail.
But it has no record of the person we have yet to become.
• Toby Moore is a Shaw Local News Network columnist, star of the Emmy-nominated film “A Separate Peace” and CEO of CubeStream Inc. He can be reached at feedback@shawmedia.com.
Click Here For The Original Source.
