Proofs, Diagnoses and Leaked Keys: AI Starts Checking the Work

Yesterday's biggest tech stories share one theme: verification. Claude wrote machine-checked proofs for real open-source code, Microsoft's diagnostic agents out-scored doctors in a Science study, a developer cloud told every customer to rotate their secrets, Ethereum paused an upgrade test to double-check gas rules, and DeepSeek raised a war chest for open models.

🧮 Claude formally verifies 123 open-source libraries

Anthropic's Frontier Red Team, working with Theorem, used Claude to prove that 239 OSS-Fuzz entry points across 123 open-source libraries never crash on any input under 64 MB, checked at the RISC-V assembly level by the Rocq proof kernel. The proofs total about 180 million lines, and along the way Claude found and patched bugs that had survived years of fuzzing. The guarantees hold only under stated assumptions (single-threaded static RISC-V Linux binaries with a fixed memory budget).

Why it matters: formal verification used to take expert-years. If AI makes it routine, "tested" can start becoming "proven."

Source: Anthropic (https://www.anthropic.com/research/formal-verification)

🩺 Microsoft's MAI-DxO beats generalist doctors on hard cases

In a study published in Science, Microsoft's Diagnostic Orchestrator, a panel of AI agents that ask questions, order tests and debate a diagnosis, reached 83% accuracy on 304 complex NEJM cases versus 36% for generalist physicians using reference tools, and 80% versus 47% on emergency-department cases, often with cheaper workups. Caveats: the benchmark is text-only, compares against generalists rather than specialists, and TechCrunch notes much of the gain comes from the underlying models.

Why it matters: orchestrating several agents that challenge each other is becoming a practical pattern, not just a demo.

Sources: Microsoft AI (https://microsoft.ai/news/building-medical-superintelligence/), TechCrunch (https://techcrunch.com/2026/10/22/microsofts-mustafa-suleyman-says-its-ai-outperforms-doctors-on-hard-cases-as-he-chases-medical-superintelligence/)

🔑 Railway tells "all users" to rotate their keys after a breach

Developer platform Railway confirmed attackers may have accessed encrypted environment variables and, for some, obtained the decrypted contents. It briefly paused builds and deployments and urged every customer to revoke and rotate API and cloud keys, saying the breach was likely part of a broader attack. A post-mortem is promised.

Why it matters: environment variables are the keys to your databases and third-party services. Treat a platform breach as your breach.

Source: TechCrunch (https://techcrunch.com/2026/10/22/developer-platform-railway-urges-all-users-to-revoke-secret-keys-after-cyberattack/)

⛓️ Ethereum postpones the Glamsterdam test fork

The Ethereum Foundation delayed Glamsterdam's activation on the Hoodi testnet, planned for Oct. 26, after ChainSecurity warned that gas changes in EIP-8037 and EIP-8038 could change how existing smart contracts handle failed calls. No new test date or mainnet date has been set, and no live exploit has been reported.

Why it matters: this is the safety process working. Shipping late beats breaking contracts that hold real money.

Sources: Ethereum Foundation (https://blog.ethereum.org/2026/10/22/glamsterdam-hoodi-delay), CoinDesk (https://www.coindesk.com/tech/2026/10/23/ethereum-delays-glamsterdam-test-after-security-devs-flag-risk-to-existing-contracts)

💰 DeepSeek reportedly raises about $15 billion

Bloomberg reports DeepSeek raised roughly 100 billion yuan (about $15 billion), double its original target, with CATL and Tencent among the largest backers. Reuters says the round was still being finalized and details could change; the money is aimed at data centers and open models ahead of a possible listing.

Why it matters: well-funded open-weight models keep pressure on prices for everyone building with AI.

Sources: Bloomberg (https://www.bloomberg.com/news/articles/2026-10-22/deepseek-raises-15-billion-in-challenge-to-leading-us-labs), Reuters via StreetInsider (https://www.streetinsider.com/Reuters/DeepSeek+wrapping+up+%2415+billion+fundraising+to+fuel+AI+race/27217623.html)

━━━━━━━━━━━━━━━━━━━━

𝘖𝘶𝘳 𝘵𝘢𝘬𝘦 — 1𝘋𝘢𝘺 𝘈𝘨𝘦𝘯𝘤𝘺

The winners of this AI cycle won't be the teams that generate the most code. They'll be the ones that can prove it works and keep it safe. That's why, when we build products like Yvion, every agent output gets checked by something outside the agent: tests, reviews, monitors, and secrets that rotate without drama. Let AI do the heavy lifting, then make it show its work.