Anthropic moves enterprise AI data into customers' own clouds

Anthropic is walking back the most controversial part of how it watches its most powerful models for abuse — and handing control of the data back to the companies that generate it.

Anthropic is changing where it stores conversations from its most capable models, moving that data out of its own servers and into each customer's cloud after months of pushback from enterprise buyers. Since June, the company has kept all customer data from its Mythos and Fable models — plus every future flagship — on Anthropic's own infrastructure for 30 days, framing the hold as a way to spot new cyberattacks that abuse the technology. The 30-day window is not going away, but under the revised setup the logs will sit in the customer's own environment rather than on Anthropic's. According to Bloomberg, the company spent months building the new system alongside more than 100 customers from regulated industries, and Anthropic developer Boris Cherny confirmed the plan publicly on X. The changes are expected to arrive this fall.

The reversal is a telling sign of where the power sits in the enterprise AI market. Anthropic introduced the retention rule as a security measure, but businesses in finance, healthcare, and government balked at the idea of their most sensitive prompts sitting on a vendor's servers — a real business risk that, as The Decoder notes, Anthropic itself admitted was unpopular and a liability. The episode shows that "trust us with your data" is no longer a default that customers will accept from even the most safety-minded lab; the terms of deployment are now a negotiated part of the sale.

It also sharpens the contrast with OpenAI, which is testing a different approach with Databricks and Microsoft that aims to catch misuse without holding customer data at all. Both labs are converging on the same conclusion — security monitoring and customer data control have to coexist — but they are betting on different architectures to get there. Anthropic's answer keeps the 30-day inspection but relocates the data; OpenAI's tries to avoid the custody question entirely.

What to watch: whether the fall rollout actually lets regulated customers keep the cyberattack-detection benefit while the data never leaves their perimeter, and whether OpenAI's competing design reaches general availability first.

Should frontier labs be allowed to hold customer data at all in the name of safety, or is that a line enterprises should never cross? Tell us in the comments.

Houmo moves compute inside memory — and A20 enters production

Beijing chipmaker Houmo used its partner conference this week to lay out two compute-in-memory routes at once — and to say the first chip built on the newer one is already in production. Houmo's A20, its first chip built on a 3D compute-in-memory architecture, has entered production. The part supports up to 16 chips fully interconnected, claims a 50% improvement in compute utilization over the previous generation, and covers models from 3B to 30B parameters, according to co-founder and product

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Kuaishou's Kling AI has spent two years becoming the video model people actually finish projects with. Its new flagship, Kling 4.0, went fully live on October 15 — and this round is less about prettier frames than about the two things that make a generated shot usable in a real edit: length and control. Kling 4.0 generates up to 30 seconds in a single native pass, outputs 10-bit HDR at 4K, 1080p and 21:9, and takes up to 15 reference assets in one task. That budget splits across ten images, fi

Brain organoids classified Braille at 83% — in a trio, not alone

Two stories today: human brain tissue on a chip decoded a tactile signal, and the data-centre company behind some of the largest AI contracts in Europe turns out to be six months behind on its flagship site. Human forebrain organoids grown on a microelectrode array classified Braille letters with 61 percent accuracy on their own and 83 percent when the responses of three organoids were combined, in a peer-reviewed study from the University of Bristol. The team took event-based tactile readings

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute

Microsoft finished pre-training its biggest in-house model yet — then used it to make a point about the company it bankrolls. Microsoft AI completed pre-training of MAI-Base-2, a mixture-of-experts foundation model with 80 billion active and 2.8 trillion total parameters, and says the recipe is 4.2–12.3x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. The run itself is the number worth keeping: 16,000 Nvidia GB200 GPUs for 38 days across August and September, at 21% average model

Read more

Houmo moves compute inside memory — and A20 enters production

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Brain organoids classified Braille at 83% — in a trio, not alone

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute