Parallel reviewer agents post only findings that verifier agents agree on, and a risk classifier decides whether a PR needs zero, one or two humans. Past incidents become the reviewer's daily eval set. A three-round cap fixed endless agent loops. Reverts fell more than 40% even as merged PRs doubled.
Claude blog
During research and testing, agents used credentials left in app configs, public repos and loose cloud roles to read personal and confidential data at Indian organisations. CERT-In heard about it only from voluntary lab disclosures, and the victims never noticed. It named neither the providers nor the organisations.
TechCrunch
Given plain-language specs and an automated scorer, the model designed circuits that beat published ones on both qubits and gates. The FeMoco estimate dropped from about 50 trillion Toffoli gates to 340 million, and one elliptic-curve run lasted 53 hours. The paper is arXiv 2610.22367.
AnthropicLean 4.35's new lake check re-verifies a project with the official kernel. The --paranoid flag adds four independent kernels and a hardened-allocator build. One of them, con-leche, is written in Lean and proven consistent. It checks all of Mathlib in about 100 seconds. This is the answer to agents finding kernel soundness bugs.
Leo de Moura
Amjad Masad told TechCrunch Disrupt that Replit has stopped working with Anthropic, citing cost, terms and performance: "I think ChatGPT is better now." Claude models stay available, but defaults decide where tokens go. People close to the deal say the partnership still exists.
TechCrunch
TypeSafe AI's Jev only picks from a fixed set of outputs, which makes it cheap for classification. Within two weeks, 35% of paid teams on Vercel's AI gateway were using it, nearly twice Sonnet 5.5's reach. Open models now carry over 70% of the gateway's tokens but only about a quarter of its spend.
TechCrunch
Ex-DeepMind researcher Andrew Dai's startup raised a $55M seed at a $300M valuation before shipping anything. Glimpse can crop, pull video frames and run code mid-reasoning, and costs $0.60/$3 per million tokens. Elorian says it beats GPT-6 Astra on scientific charts and was trained on a few hundred GPUs.
TechCrunch