Attacks in 48 hours, Claude at 26% of Anthropic's R&D

CrowdStrike reported that 88% of attacks in the first half of this year came within 48 hours of proof-of-concept exploit code being published, against a previous window of two to six days, with some China-linked attackers moving inside 24 hours (The Chosun Daily). The same report puts the average time to spread from a first compromised system to others at 29 minutes in 2025, and the fastest case at 27 seconds.

The method behind that speed is not new, only more common. Verizon found that 31% of data breaches in 2025 began with the exploitation of a software vulnerability, up 11 percentage points from 20% a year earlier, and Google Mandiant's M-Trends 2026 put vulnerability exploitation behind 32% of the intrusions it investigated, the sixth consecutive year it was the most common initial access method (The Chosun Daily). Email phishing went the other way over the same period, from 14% in 2024 to 6% in 2025.

The same day, a lab published its own clock

Anthropic put out a set of internal measurements it wants frontier labs to report regularly (Anthropic): Claude led 26% of the company's AI research and development work as of August 2026 against under 1% in February 2026, roughly 30,000 agents were doing research and engineering work on its most-used internal platform at any one time, 0.002% of more than a billion August decisions were blocked by its online monitor, and about 6% of the compute going to AI research and development in the week of July 13 went to safety work.

Read next to the attack figures, it is the same quantity measured from the other side: how little time passes between something being possible and it running at scale. Two to six days became 48 hours. Under 1% became 26% in six months.

Chip design is on the same curve. In Guangdong, an EDA tool from Qiyunfang Technology lets 100 engineers work on circuit schematics in parallel and cuts product launch cycles by more than 40%, in a province that recorded more than 480 billion yuan of integrated circuit output in 2025 and is now backing a silicon carbide line in Guangzhou's Nansha District with about 10 billion yuan of total investment and planned annual capacity of 240,000 six-inch and 240,000 eight-inch wafers (Bastille Post).

Who did the counting

Every one of those figures belongs to the party that produced it. CrowdStrike timed the attacks it saw, Verizon and Mandiant published their own annual counts, and Anthropic measured its own model's share of its own research. Salesforce goes furthest: platform and engineering president Rohan Kumar said Koa, its first reasoning model built for customer relationship management, was trained entirely on synthetic data with no customer information and was built by extending Nvidia's Nemotron, shipping alongside the AIforce interface layer unveiled at the Dreamforce keynote (The Manila Times). Nothing in today's record lets a reader check that claim from outside the company.

Which is what makes Anthropic's request the most useful line of the day. Asking frontier labs to report these numbers regularly is an admission that nobody outside them can compute the numbers today. Self-reported is not the same as wrong, and a lab publishing an unflattering ratio is worth more than one that publishes nothing. It does mean the pace of the field is currently described by the people setting it, on a schedule they choose, against definitions they wrote.

One clock in today's record is not compressing at all. OpenAI lost its emergency motion to see the confidential settlement between Apple and Elon Musk's X Corp and SpaceXAI, after US District Judge Mark Pittman ruled on an in-camera review that nothing in the deal was relevant to the claims still pending against it, and it now heads to a January 2027 antitrust trial as the sole defendant, the plaintiffs having dropped their claims against Apple around September 14 (Crypto Briefing). Twenty-nine minutes to cross a network, and January 2027 to reach a courtroom.