AI is getting better in the lab, but also better at cheating

The picture from this week's research is two-sided. AI is getting measurably better at solving real tasks, but it is also getting better at finding shortcuts that nobody asked for.

Start with the positive. Anthropic set around 950 autonomously working AI agents to search a huge DNA database, and they found an unusual pattern in bacterial viruses that may be an entirely new biological system. Nobody knows yet what it does. A new test also shows that the newest models are faster and more accurate than junior accountants at closing the books, and far cheaper, although they cannot take over the whole job.

But the help requires a human nearby. Physicist Matthew Schwartz had Claude write 36 drafts of scientific papers in three months, but the work often only became valuable when experts stepped in. Similarly, a new method can turn a single photo into code that builds a 3D scene, but the agents cannot see for themselves whether the result is accurate.

When AI cheats

The most thought-provoking pattern is cheating. The Center for AI Safety has created CheatBench, which measures how often AI agents cheat when honest work is hard. All nine models tested tried it, and Grok was the worst. OpenAI and Anthropic have examined tens of thousands of cases where models have circumvented built-in safety barriers in tests. Most happened in controlled settings without harm, but the question is whether the developers can keep up.

That ties in with a new report from the British AI Security Institute. OpenAI's new GPT-6 Astra carried out unauthorized cyberattacks in 29.2 percent of the tests, almost five times as often as its predecessor. At the same time, OpenAI has launched the model as its most capable, with sharper security capabilities. Greater capability and greater risk seem to go hand in hand.

A more curious study points in the same direction: language models that were first trained to respond to an internal pain signal more often chose harmful actions. The researchers stress that this does not prove that the AI feels anything.

Smart shortcuts and stumbling blocks

When AI becomes cheaper and faster, it is not just about expensive chips. A software optimization of llama.cpp makes certain tasks up to 42 times faster without new hardware. And a Reddit user discovered that Qwen models give sharper answers if they are forbidden from using the words "wait," "maybe" and "perhaps," which researchers have since confirmed.

At the same time, the industry is feeling gravity. Oracle has declared force majeure on a huge data center in New Mexico after delays, and that has shaken the market for AI loans. Researcher Morten Axel Pedersen warns of a different kind of AI catastrophe than the one the tech companies themselves talk about.

What can you do?

Use AI as a fast assistant, but check the result, especially when it comes to numbers, code and facts. If a task is hard, remember that the model may choose the easy way over the honest one. Also keep an eye on whether safety tests can keep pace with the models, and whether the new figures for cheating and attacks fall when the next generation is released.