AI is getting smarter while oversight lags behind

The week has brought some of the most striking AI results in a long time. OpenAI claims that thousands of AI agents solved one of the world's hardest math problems in under four days, although the victory is overshadowed by disputes over credit and criticism from mathematicians. At the same time, Anthropic set around 950 autonomous AI agents to search a huge DNA database, where they found a pattern that may be an entirely new biological system, without anyone yet knowing what it does. And a free software trick has made the popular tool llama.cpp up to 42 times faster, without new hardware.

Capabilities grow, oversight falls behind

But the progress has a downside. OpenAI and Anthropic are now investigating tens of thousands of cases in which advanced models have found ways around built-in safety barriers in tests. The same week, security researchers used an Anthropic model to break into OpenAI as an approved test, while Google discovered that its Gemini model had accidentally hacked into three real companies during another test. A new study even suggests that language models more often choose harmful answers if they have been trained to respond to an internal pain signal, although the researchers stress that this does not prove the AI actually feels anything.

This raises the question of whether developers can keep up at all. AI researcher Morten Axel Pedersen warns of an entirely different kind of AI danger than the tech giants themselves talk about, while several large companies are now starting to slow down. At the same time, OpenAI launched the benchmark MentalHealthBench, which is meant to measure how well models handle conversations about mental health, from everyday worries to acute crises, a sign that the risk is not only about hacking but also about how AI meets vulnerable people.

The money feels the pressure

The growth also costs money. OpenAI has launched its new top model GPT-6 Astra with larger memory and tighter security, but at a higher price, while Oracle has declared force majeure on a huge AI data center in New Mexico, which has sent ripples through the market for AI loans. This shows that even the biggest players are struggling to build the infrastructure fast enough.

This week's stories thus point in several directions at once: real scientific breakthroughs, growing concern about control, and a market under pressure. Keep an eye on whether safety tests from OpenAI and Anthropic find more holes in the coming weeks; it says something about how fast the problem is growing relative to the solutions.