GPT-5.6-Cyber, out August 10, complies with 95 percent of advanced security prompts that standard Sol declines, per OpenAI's announcement. Daybreak now splits into a Blue tier for defenders with no ID check and a Red offensive-simulation tier that requires a verified organizational credential. The gate is the real story: capability this sharp only ships behind identity. Read our take →

Anthropic locked a $9.1 billion, 20-year compute deal with Riot Platforms for 191 MW in Rockdale, Texas, per QZ and Bloomberg. Its third large compute commitment in three weeks, and Riot stock jumped 25 percent after hours. Same week, Nvidia signed MOUs with six asset managers including Apollo, Blackstone, and KKR to mobilize more than $500 billion in third-party capital, treating GPU compute as collateral the way lenders treat real estate. Compute is now an asset class you finance, not just a bill you pay. Anthropic and Riot → · the Nvidia pact →

Corma raised a $60 million seed from Sequoia, Khosla, and Coatue to build a foundation model trained only on defensive security. The number behind the check: Corma's simulations found AI attackers succeeded 88 percent of the time while AI defenders caught just 12 percent, per Fortune. That gap is the whole thesis, and it is the same gap OpenAI's new cyber model is chasing. Our writeup →

Since August 2, every new Claude model embeds machine-readable watermarks in generated text and C2PA provenance metadata in generated images, across Claude.ai, the API, Claude Code, and the AWS, Google Cloud, and Microsoft Foundry deployments. Run Claude output through a publishing or LLM pipeline and your content now carries embedded attribution, so decide where that matters before it surprises you. Read our take →

Wire a local model like Muse Glimmer to the outside world without writing glue code. Make builds the multi-step automations that hand your agent real API actions from a visual canvas. Try Make →

Feed clean, LLM-ready web content to a local model with no cloud API in the loop. Firecrawl turns any URL into structured markdown your agent can actually parse. Try Firecrawl →

Rank agent platforms on harness architecture, evals, and lock-in with a scorecard you can hand to procurement. Read it →

Pulled from the public record of real enterprise rollouts, not vendor decks. Read it →

Draw the data boundaries before security asks where the training data went. Read it →