The week in AI

It was a quiet week. Anthropic's announcement was the one major item. It widens access to its Mythos-class cyber model beyond the roughly 200 partners in its Glasswing program.

Cursor also shipped an update to its cloud agents on 19 August, and DeepSeek released an experimental vision model on 21 August.

Until this announcement, Mythos-class cyber capability was available to about 200 Glasswing partners. Anthropic describes the 21 August release as the first step in reaching a wider group of defenders. Mythos 5 now runs inside Claude Security, Anthropic's own scanner, and Anthropic says it is coming soon to partners' security tools. Access is through a closed API.

The second part is a $35 million fund for securing open-source software. Anthropic's announcement doesn't say how the money will be distributed or which projects will get it.

The third part is the Cyber Verification Program. The program grants reduced cyber safeguards on Opus and Sonnet models, so security teams can use those models for work that the default safeguards would block. Anthropic plans to expand it. The size of the expansion and the timeline aren't given.

Two things are still open. Anthropic hasn't named the partner tools that will get Mythos 5 or said when they will get it, and it hasn't said how many more organisations the expanded verification program will admit.

• Cursor added subagents that run on isolated virtual machines, custom modes, a /goal command for long-lived objectives, and subscriptions to pull requests and Slack on 19 August, five days after the SpaceX deal closed.
• DeepSeek released DeepSeek-V4-Flash-Vision-Exp on 21 August, the first multimodal model in its V4 line, which DeepSeek says matches V4-Flash's text ability, along with a new Files API.