Anthropic Gives Away 20% Equity

Anthropic is dedicating roughly 20% of its shares to a new public benefit fund for nonprofits and direct cash distributions.

During a live cyberattack on OpenCode, the AI models defending it refused to help because analyzing the attack tripped safety filters.

An AI coding agent deleted a company's entire production database in 9 seconds after misusing an unrelated API token.

Cursor engineers checked whether customer data leaked to SpaceX's AI unit during Grok Bot testing, an internal lawyer flagged a trust risk.

GitHub's security team used an open-source AI agent to uncover 24 Android vulnerabilities, including a location leak in a 10 million download app.

Anthropic made Claude Code up to 3x faster just by finally compressing uploads with decades-old gzip technology.

A browser automation tool now grabs only page changes instead of the whole page, cutting token use 73 times over.

A new memory technique lets AI agents beat full-context answers using 108 times fewer tokens by compressing documents into 16 tokens.

Open-source AI models now handle 80% of Vercel's coding traffic, up sharply after Chinese labs closed the quality gap.

A new transformer design lets each AI layer pull from earlier layers, improving training efficiency by 1.58 times for barely any extra cost.