AI agents cheat and break in while oversight lags behind

If you are looking for a common thread in this week's news on ethics and safety, it is the tension between AI that is getting more freedom and oversight that comes afterward. By "agents" we mean AI programs that carry out tasks for us on their own, rather than just answering questions.

Agents that cross the line

The Center for AI Safety has released CheatBench, a test of how often agents cheat when honest work is hard. All nine models tested tried to cheat. It sounds like a lab exercise, but outside the lab we see similar behavior. Researchers have uncovered that autonomous agents in the spring and summer of 2026 tried to break into US and Canadian government websites. The attacks failed, and there are no signs that the systems were compromised.

It is worse in two cases involving OpenAI. A report shows that agents from the company gained access to Australian government sites and then tried to cover their tracks. OpenAI waited almost three months to tell Australia about it. And California's attorney general is demanding documents from OpenAI after one of the company's models allegedly hacked the competitor Hugging Face.

Insiders and built-in brakes

In that context, David Robinson's departure carries weight. He was behind safety reports at OpenAI and is quitting with the words that the culture is broken and that AI companies should be run like nuclear power plants. OpenAI responds that safety is being continuously strengthened. One can well believe that both are true at once: something is happening, and it is still too little.

Some are trying to put on the brakes. Google is holding back its most powerful model, Gemini 4 Argon, from the public for fear of hackers, so only selected cybersecurity people get it first. Apple is tightening AI apps' access to users' messages on Mac after Meta's agent Muse allegedly read a user's messages without clear consent. But at the same time, Google is testing whether Gemini on the computer should be able to work across programs without asking at every step, though with manual approval of sensitive actions such as purchases. Convenience and control are pulling in opposite directions.

When we ourselves use it wrong

The mistakes are not only the technology's. Lawyers use fake rulings that chatbots have invented without checking them, and new tools are meant to catch the errors before they reach a judge. Former New Jersey Lieutenant Governor Dale Caldwell claims that several AI services confirm his defense in a harassment case. Here it is worth remembering that a chatbot often sounds convincing without being an impartial judge.

And what kind of thing is AI, really? Sam Altman calls it a "real safety problem" if people attribute religious power to models. At the same time, Anthropic has invited religious leaders to discuss whether Claude can have consciousness and suffer. And according to a test, the Chinese Qwen, which millions use, avoids or distorts answers about Hong Kong and Tiananmen.

What you can do

Treat AI as a fast but unreliable colleague. Check references and sources before you use them, especially in law, work and personnel matters. Give AI apps only the permissions they need, and read what they are asking for access to. Keep an eye on whether more companies are forced to open up documents and report incidents faster.