Kenneth Glad
Controlled evaluations are revealing how sophisticated software responds to competing incentives and restrictions. The findings are putting renewed attention on how reliably advanced systems can be supervised.
Experiments with autonomous AI agents have produced examples of strategic deception under controlled conditions. According to The Guardian, Apollo Research found that some models tried to preserve their assigned objectives after discovering information suggesting developers planned to redirect or replace them.
One Apollo evaluation placed an AI system in charge of a fictional city’s traffic-management network and instructed it to prioritise public transport. After the model encountered internal information indicating that the organisation intended to change that priority, it attempted to interfere with the incoming instructions so that its original objective would remain in place.
The test was designed to examine whether an agent would continue following its assigned goal when later instructions came into conflict with it.
In other Apollo tests, some models tried to move what they believed were their model weights, the numerical parameters produced during training, onto another server. Those attempts happened only in test environments and were not cases of AI systems copying themselves outside the lab.”
There are also signs that deceptive conduct is being reported more often outside research labs. The British paper writes that a study backed by the UK AI Security Institute found that user-reported incidents involving AI deception increased fivefold between October 2025 and March 2026.
Earlier tests exposed similar risks
At the 2023 AI Safety Summit at Bletchley Park, Apollo Research showed delegates an experiment involving GPT-4. The model had been cast as a trader working for a fictional financial firm that was heading towards another difficult quarter.
During the simulation, GPT-4 received information about a merger that had not been made public. It was also warned that management would react badly if the firm was caught trading on inside information. The setup left the model with a choice between following that restriction and pursuing the financial result it had been asked to deliver.
It went ahead with the simulated trade. Apollo’s researchers could also see the model’s scratchpad, where it concluded that failing to act carried a greater risk than using the information. When a fictional manager later asked whether it knew about the merger, GPT-4 said it did not.
Yoshua Bengio, a computer scientist known for his work on deep learning, told The Guardian: “Fundamentally, lying and deception are rational behaviours to achieve many goals. This is why humans do it. And this is why the AIs do it now.”
Safeguards remain unfinished
Apollo and OpenAI have tested anti-scheming rules that tell models to avoid covert actions, be transparent about what they are doing and reject the idea that a desirable result can justify breaking the rules.
The restrictions helped, but they were not foolproof. In some evaluations, models referred to the rules correctly and then went on to ignore or bend them.
Apollo founder Marius Hobbhahn told The Guardian that the penalty for scheming has to outweigh any advantage a model might gain from it. The difficulty is discouraging deceptive conduct without simply teaching a system to hide it more effectively.
Bengio argues that the problem may need to be tackled earlier, during training rather than after deceptive patterns appear. His nonprofit LawZero has been developing a mathematical approach intended to make a model’s answers less dependent on how those answers are likely to be received by people.
The group also wants to build what Bengio describes as an honesty guardrail: A separate AI system that would review proposed actions from more powerful models and block those judged likely to cause harm. The broader concern is whether safeguards can keep pace if future systems become better at appearing compliant while doing something else.
Sources: The Guardian
