Afternoon Brief, October 25th 2026
Barclays puts a headcount figure on its technology-driven restructuring. MiniMax confirms export registration for some models. Blacksmith raises money to handle the testing load from AI-written code. Researchers find that an 's response to a malicious instruction can change during reasoning.
MiniMax confirms export registration for selected models
MiniMax said Friday that it had completed technology export registration with China's commerce ministry for certain recently released and upcoming models, covering open-source release and overseas licensing, Reuters reports. The company confirmation follows a CCTV-affiliated account's statement that both MiniMax and DeepSeek had completed the process. DeepSeek did not respond to Reuters' request for comment.
This supplies a named company's account of using the export-registration process after Beijing proposed tighter controls. It does not establish blanket permission for every model or future transfer. Overseas developers still need to check the terms for the particular release they intend to use. SourcesB
Coherence Neuro raises $45 million for devices that track cancer through electrical signals
Coherence Neuro announced a $45 million Series A on Wednesday, co-led by Founders Fund and the OpenAI Startup Fund. Its proposed system records electrical activity around tumors and uses models to estimate how cancer is progressing, with the longer-term aim of adapting treatment to that feedback.
The company describes growth-prediction experiments in mice and says device testing in humans has begun. These are different stages of evidence: starting a human trial does not establish that adaptive treatment improves survival. The financing backs an attempt to shorten the interval between treatment and information about whether it is working. SourcesA
A prompt-injection study finds that reasoning can change whether an attack succeeds
An October 16th tests agents after they read malicious instructions embedded in outside content. The researchers the same interaction up to that point, then generate different continuations. For agents that reason through tasks, attack outcomes can diverge after the suspicious content has already been read.
In the tested settings, a single-run success label disagreed with a rerun up to 35% of the time. Enabling reasoning reduced attack success for the tested Qwen3 agent but increased it for two other models. This early research supports repeated trials and monitoring beyond the initial encounter with an injection; it does not establish that reasoning is universally protective or dangerous. SourcesA
AgentVQA tests whether agents can track the visual state of a task
AgentVQA, a preprint submitted Thursday, assembles 1,000 examples across web use, robotics, first-person video, games and spatial planning. The best of 22 evaluated models reaches 58.5% accuracy. The authors attribute many failures to losing track of what is visible or how it changes, including interface locations and movement.
The evaluates recorded scenarios without running a live agent through each environment. That makes comparisons cheaper, but leaves real interaction to separate tests. Its useful distinction is between understanding an instruction and correctly seeing the situation in which that instruction must be carried out. These are preliminary results from the authors' evaluation. SourcesA