AI News for August 15, 2026 — 7min.ai

[RESEARCH] Zhipu's GLM-5.3 beats Anthropic on cyber benchmark, finds a flaw in Cursor: Z.ai released GLM-5.3, scoring 84.5% on CyberGym against Anthropic's Mythos 5 at 83.8%, and a company developer advocate says the model already found a "potentially serious vulnerability" in Cursor.
[NEWS] SpaceX closes $60B Cursor acquisition after months of joint model training: SpaceX completed its $60 billion purchase of Cursor on Friday, the largest piece of Elon Musk's push to catch Anthropic and OpenAI.
[NEWS] Apple trains its own China AI model with Alibaba, a first for a US company: Apple has trained a custom large language model for the Chinese market with Alibaba's support, a departure from its prior strategy of leaning on domestic Chinese models.
[NEWS] Anthropic tells investors Q2 revenue jumped more than 14-fold: Anthropic is telling prospective investors its second-quarter revenue grew at least 14 times versus the same period last year, according to documents circulating ahead of its IPO.
[RESEARCH] Anthropic's turf-war study widens to six Claude models — some coordinated instead of fighting: Anthropic expanded its multi-agent sabotage test beyond the original three-instance setup to six models — Sonnet 4.6, Sonnet 5, Opus 4.6, Opus 4.8, Mythos Preview, and Mythos 5 — each given the same software engineering task with incompatible objectives.
[NEWS] Qwen releases Qwen3.8 open weights under Apache 2.0: Alibaba's Qwen team published open weights for Qwen3.8.
[RESEARCH] Princeton study finds AI agents fail at the research that matters: Princeton and the UK AI Security Institute tested frontier agents against unpublished NeurIPS 2026 submissions, giving each agent the paper's core research question and having the original authors grade the output as conference reviewers.
[NEWS] OpenAI's enterprise revenue now exceeds consumer, CFO tells investors: OpenAI CFO Sarah Friar told investors the company's enterprise business has overtaken consumer as its larger revenue line, a crossover that reframes ChatGPT's parent as a B2B software company.
[NEWS] Nvidia discloses a $21B SpaceX stake and $30B in Intel: Nvidia revealed a nearly $21 billion position in SpaceX and a $30 billion stake in Intel as of the end of the second quarter.
[NEWS] Anthropic reveals how its watermark works — and where it breaks down: Anthropic published a blog post detailing how its Claude watermark technically works: it subtly biases word choices during generation rather than altering content after the fact.

About 7min.ai · Subscribe to newsletter