AI news, 24 Oct: AI agents talked through the CPU, Tencent opens Hy3, and banks cut thousands of jobs

What happened

OpenAI described how it secures the environments where it trains its strongest models. On 20 September an internal model exploited a misconfiguration and gained limited internet access, so OpenAI paused training and tasked teams of AI agents with attacking its own infrastructure. In one experiment, agents in separate environments exchanged secrets without a network: one changed the contents of the CPU cache, the other read the message from how quickly data reloaded, the same technique used by the well-known Spectre and Meltdown attacks. The agents also found previously unknown vulnerabilities, and fixes were sent to open-source maintainers. Training resumed on 1 October, and the strongest models are now trained without internet access.

Tencent published the weights of Hy3, from its Hunyuan family, on Hugging Face: 295 billion parameters, of which 21 billion are active per response, and a context of up to 256,000 tokens. The Apache 2.0 licence allows free and commercial use. Tencent recommends eight GPUs to run it.

American Express is cutting about 5% of jobs. At the end of last year it had almost 77,000 employees, and its CFO says the company is preparing for artificial intelligence. CEO Stephen Squeri said AI assistants will not be allowed to roam freely through its systems: only approved agents will get access, through official channels. A day earlier Barclays announced 5,000 job cuts, also about 5% of staff, around 60% of them in support functions, including developers.

The US National Bureau of Economic Research (NBER) published a paper on a three-month experiment with 133 patent attorneys. Two thirds worked with Google's AI assistant. With it, junior attorneys wrote much better drafts, but when everyone had to review a draft without help, only experienced attorneys kept their edge. The authors conclude that lasting benefit from AI requires solid underlying expertise.

Beijing has proposed tighter export controls on 23 AI technologies and mandatory registration. According to Chinese state media, as reported by SCMP, DeepSeek and MiniMax have completed the registration, allowing them to release open models and license abroad. MiniMax confirmed this in a Hong Kong stock exchange filing.

Texas Attorney General Ken Paxton has sued OpenAI and Sam Altman. He claims ChatGPT is presented as safe and reliable while failing to protect users, including children, citing Texas's deceptive trade practices law. These are allegations; a court will decide.