Quick Hits — August 14, 2026
An evening catch-up: Anthropic nudges its misalignment risk up a notch, Thrive's Joshua Kushner tells Silicon Valley to cool it on AI, and Z.ai opens its vulnerability books.
Anthropic raised its misalignment risk estimate from "very low" to "low" and says it has no plans to release its stronger internal model, "Model 2." In its August risk report, the company cited recent cybersecurity incidents for the downgrade and flagged "signs of acceleration" in models' ability to conduct automated R&D — a capability that could speed up technical progress or be misused. The report describes Model 2, used heavily for internal coding and agentic work, as a "noticeable improvement" over previous models, but says releasing it externally isn't on the table. The take: the most striking line is Anthropic admitting its own task-based evaluations "no longer capture increases in models' capabilities" — a frontier lab openly saying it can't fully measure its models, right as OpenAI pauses Astra over cyber-risk concerns.
Joshua Kushner used Thrive Capital's first-ever investor letter to chide Silicon Valley VCs over AI euphoria — "It would also be a grave error in our minds to let excitement weaken our investment discipline." In the letter, leaked to Bloomberg, the OpenAI backer argues Thrive's concentrated approach (about 90 percent of capital in its top 15 investments per fund, no spray-and-pray) beats the "outlier" philosophy of making many bets and hoping a few hit. The take: it's easy to read this as a rich man's humblebrag, but it lands at a moment when AI dealmaking is running on hype — and Thrive has as much on the line in AI as anyone.
Z.ai, the international arm of Chinese lab Zhipu, opened a public security disclosure ledger that catalogs more than 2,400 vulnerabilities — with just 53 disclosed so far. The site quietly went live this week and is now the clearest window yet into how the GLM maker triages model and product security findings: 2,383 items still sit unpublished in the backlog, with the ledger updated in both Chinese and English. The take: at a moment when Chinese labs face scrutiny over cyber-capable models, publishing the raw ledger — backlogs included — is a transparency move most Western labs wouldn't make.
