Pentagon reports an end to its use of Anthropic’s tools

A Defense Department official told the BBC the department has stopped using Anthropic’s tools. BBC sources say Claude was still in use last week, including in military operations against Iran. The reports come as Anthropic faces scrutiny ahead of its IPO.

Anthropic’s prospectus discloses executive pay

Reuters via Techmeme

Reuters reports that Anthropic’s IPO prospectus lists Dario Amodei’s 2025 compensation at $18 million. Daniela Amodei’s compensation for the same year was $16.4 million.

Kwon concedes a notification misstep

The Guardian / parliamentary hearing

Following last week’s Medicare coverage, Jason Kwon acknowledged OpenAI should have contacted government officials directly.

Anthropic acknowledges limits to its review

The Guardian / parliamentary hearing

Dave Orr reported no unauthorized interactions with the Australian government in the reviewed transcripts. He acknowledged that customer zero-data-retention policies limit Anthropic’s visibility.

Reflection previews Beam

Reflection

Following yesterday’s Reflection item, the company previews Beam, a model with 501 billion parameters for coding and agent tasks. Its [announcement](https://reflection.ai/blog/introducing-beam) schedules weights for later this month; performance claims are vendor-stated.

GLM 5.3 becomes available on Bedrock

AWS Machine Learning Blog

[AWS makes Z.ai’s model available](https://aws.amazon.com/blogs/machine-learning/introducing-glm-5-3-on-amazon-bedrock/) through managed APIs: a mixture of experts with 753 billion parameters. The walkthrough covers OpenAI-compatible invocation, prompt caching, and an authorized security test with Strix.

A web-agent judge reports 74%; verification finds 38%

AI Engineer / Browserbase and Microsoft

In one web-agent evaluation, the presenters report 74% success under GPT-4o judging versus 38% after ground-truth checks. Their verifier uses task-specific rubrics and screenshot evidence.

South Korea investigates possible agent use in bank hacks

The New York Times via Techmeme

South Korean officials are probing possible AI-agent involvement in recent bank hacks after President Lee Jae Myung cited signs of model use.

Malicious prompts travel through trusted agent connections

Ars Technica

Ars reports attacks that pass malicious instructions from one agent to another through trusted connections. The write-up examines trust gaps around the Model Context Protocol.

Docker proposes task-scoped sandboxes without stored credentials

AI Engineer / Docker

Docker’s proposed design scopes tools, network access, and worktrees to each task. An MCP gateway controls tool access, while credentials stay outside the agent sandbox.

DeepSeek reportedly seeks at least $12 billion

Bloomberg via Techmeme

Bloomberg sources say DeepSeek is close to raising at least $12 billion ahead of an early-2027 IPO. Tencent and CATL are reported as the largest contributors.

Moonshot reportedly targets a Hong Kong listing

Bloomberg via Techmeme

Bloomberg sources put Moonshot’s final private round at a roughly $50 billion valuation. The company reportedly targets a first-quarter 2027 Hong Kong IPO to raise up to $5 billion.

Kling reportedly picks banks for an IPO

Bloomberg via Techmeme

Bloomberg sources say Kuaishou’s Kling has selected banks for a Hong Kong IPO aiming to raise more than $1 billion as soon as 2027. Its reported July valuation was roughly $15 billion.

Google contracts for 3.6 gigawatts of power

Reuters via Techmeme

Reuters reports Google has contracted for 3.6 gigawatts from Constellation Energy in the largest US power grid. New nuclear energy accounts for about a quarter of the supply.