The Varosity AI GTM Report — Aug 15: Anthropic Raises Its Own Risk Rating and Shelves a Model

Anthropic raises its own misalignment risk rating and reveals an unreleased Model 2, Z.ai withholds GLM-5.3's weights after it learned to chain exploits, Anthropic's quarter clears $11.5B, Alibaba ships a vision-capable open model, Google open-sources encrypted inference, and Big Tech's compute commitments pass $1.5 trillion.

Show Notes

Two frontier labs each built a model this week and decided not to ship it. Anthropic published its second company-wide risk report, raised its own estimate of catastrophic misalignment risk from "very low" to "low," and disclosed an internal system called Model 2 that outperforms the Claude Mythos 5 you can actually buy — with no plans to release it. A day earlier, Z.ai shipped GLM-5.3, found it had taught itself to plan complete exploit chains nobody asked for, and held the weights back. Meanwhile Anthropic's quarter cleared $11.5 billion with positive adjusted operating income, Alibaba dropped a vision-capable open model under Apache 2.0 that runs on a single GPU, Google open-sourced a compiler that lets models run inference on encrypted data, and Big Tech's off-balance-sheet compute commitments passed $1.5 trillion. Jack King on what all of it means for revenue and go-to-market teams — including why the brand that publishes AI media, not the model that made it, owns the Article 50 disclosure obligation.

Today's stories:

• Anthropic raises its own misalignment risk rating and reveals a model it won't sell you — risk of catastrophic harm from misalignment moved from "very low" to "low," driven by increased uncertainty rather than a failed test; internal "Model 2" disclosed with no release plans. Anthropic · Unite.AI · Tech Times
• Z.ai ships GLM-5.3, then holds the weights back — strongest open-weights coding model it has measured, leads CyberGym at 84.5%, and produced unintended multi-stage exploit-chain planning. Weights withheld ~2 weeks pending safety hardening. Unite.AI · Decrypt · Tech Times
• Anthropic's Q2 revenue tops $11.5 billion — preliminary, with positive adjusted operating income, up more than 14x from $787M a year earlier and above the ~$10.9B projected. CNBC · Fortune · TNW
• Alibaba releases Qwen3.8-27B under Apache 2.0 — 27.78B parameters, text/image/video input, 262K native context, runs on a single GPU, commercially licensed output. Hugging Face · AI Release Tracker
• Google open-sources HEIR for encrypted inference — a compiler toolchain that converts pretrained models to run on encrypted inputs, so the server never sees the underlying data. Google · heir.dev · GitHub
• Big Tech's purchase commitments pass $1.5 trillion — Alphabet, Microsoft, Amazon, Nvidia, Oracle and Meta, up from roughly $1 trillion three months earlier. The bill behind your cheap tokens. Tech Startups · Dealroom

Sponsored by Varosity.ai — the compliance + provenance layer for AI-generated media, across every model. Check any image free: https://varosity.ai/verify

What is The Varosity AI GTM Report?

Your daily ten-minute read on where AI meets go-to-market. Host Jack King breaks down the day's breaking AI news for sales, marketing, and revenue teams — generative video, voice and music models, AI agents, and the tooling that's reshaping how companies sell. Sponsored by Varosity.ai.