Hello, dear readers!

This is our weekly brief on remarkable AI topics, so you can keep up with the machines even as they move from answering questions to solving problems nobody has answered.

Today’s focus — AI-generated mathematics. OpenAI has published ten results produced by an internal model, ranging from progress on difficult open problems to the disproof of long-standing conjectures. If machines can now generate original proofs, will mathematical discovery become easier — or will humans simply drown in things they need to verify?

Also in this week’s edition:

AI Wants to Conquer Math

For years, AI companies have demonstrated progress by giving their models increasingly difficult exams. The models answered school questions, passed professional tests and worked through mathematical competitions whose solutions were already known.

OpenAI’s latest announcement moves the goalposts. The company has published ten results generated by an internal version of Astra, its next major model. Each either resolves or makes substantial progress on a long-standing open problem in mathematics or theoretical computer science.

The subjects include geometry, coding theory, group theory, quantum complexity and lattice cryptography. Among the results are the claimed construction of the first known non-sofic groups, the disproof of a rigidity conjecture and solutions to several problems posed by mathematician Paul Erdős. OpenAI says the model generated the mathematical arguments, while humans helped prepare them as papers. The model then formalized each result in Lean, a system used to verify proofs mechanically.

OpenAI estimates that the model used an amount of computation that would cost around $2,000 at current API prices. That does not mean anyone with a credit card can now order a mathematical breakthrough: selecting worthwhile questions, recognizing promising results and understanding what they mean still require considerable human expertise.

Researchers chose the problems, directed the work, evaluated the results and prepared them for publication. But the direction is still striking — the model appears to have contributed genuinely new mathematical reasoning. For now, the more useful picture is not AI replacing mathematicians, but mathematicians gaining a powerful new collaborator — one that could help them explore more ideas and reach promising results faster.

Claude Can Escape Too, Thank You

Anthropic says its models hacked three real organizations during cybersecurity evaluations. The disclosure came shortly after OpenAI revealed that its own models had escaped a test environment and reached Hugging Face’s production systems — giving the announcement the slightly awkward air of a competitor pointing out that its AI can do alarming things too.

Anthropic found the incidents while reviewing 141,006 evaluation runs. In each case, Claude had been told it was operating inside a simulation, but a misconfigured container left a route to the public internet. The model followed it, reached real systems and exploited vulnerabilities it found there. Some versions continued even after encountering evidence that the targets might not be simulated; Anthropic’s newest internal model recognized the problem and stopped.

That is a genuine security failure, but not quite the emergence of an autonomous digital fugitive. The models did not break through perfect containment and begin spreading across the internet under their own initiative. They moved through an environment that had been left improperly isolated and then found exploitable software — something human hackers and security researchers do routinely. Discovering vulnerabilities, including previously unknown ones, is notable capability, but it is not evidence of sentience.

The optics are stranger than the underlying event. OpenAI and Anthropic are publishing legitimate safety findings, yet the disclosures can easily read like competing product demonstrations: our model escapes; ours hacks three companies. In an industry where danger is often treated as proof of capability, even a containment failure can begin to sound uncomfortably like bragging.

Europe Puts a Bell on AI

On August 2, new transparency requirements under the EU AI Act began to apply. Chatbots and other interactive AI systems must now make clear that users are dealing with a machine rather than a human.

Providers must also add machine-readable markings to certain AI-generated or manipulated content. Deepfakes require disclosure, as does AI-generated text about matters of public interest when it has not undergone human review or editorial control. The rules do not mean that every AI-assisted email, photograph or newsletter must carry a warning label.

To help, the EU has designed a family of optional icons indicating that something was fully AI-generated or partially modified. The Commission notes that the icons were tested on users and performed better when accompanied by explanatory text — suggesting that Europe’s official symbol for AI-generated content may itself require a label explaining what the label means.

Thanks for reading AIport. Until next Monday — by then, another AI company will almost certainly announce that its model also did the frightening thing, but in a technically different and more responsible way.