GitHub Copilot Fusion | AI Field Notes #126

AnalysisPicking a model is no longer the developer's job inside GitHub Copilot. Fusion, which went generally available at GitHub Universe on October 28, routes each prompt to one model or several, either letting a cheap model draft while a check decides whether to escalate, or having one model write and a second critique it. GitHub claims the top Max setting scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code on Opus 5.5, at roughly half the cost per task. Billing stays per model at standard rates, so the router now decides your bill too.

AnalysisThe scorer feature teams use to grade LLM output became the way in. CVE-2026-116692, published October 27 with a severity score of 9.8 out of 10, affects MLflow (the open-source tool many teams use to track experiments and evaluate models) from 3.12.0rc0 up to 3.17.0. An attacker uploads a Python module through the artifact endpoints, then names it in a call to the scorer invoke API, and MLflow imports and runs it. Deployments with no auth configured need no credentials at all. So far the only public code is a version scanner, which usually means the clock has started.

AnalysisSixty-two critical or high-severity Chrome bugs, 13 of them sandbox escapes, came out of one Anthropic research effort. The company's Frontier Red Team said on October 27 that an unreleased internal Claude model, given initial prompting and little else, chained flaws into five working attacks: a one-click Chrome escape, a Safari same-origin bypass, a lock-screen photo grab on a Pixel 9, a Windows QUIC crash and a zero-click Linux root shell that took about 53 minutes. Everything was patched before disclosure. The Hacker News found two CVE records describe different code than Anthropic does, so some of the credit rests on Anthropic's word.

AnalysisForecasting tomorrow's electricity load or next week's orders may no longer need a GPU, if Mistral's numbers hold. The French lab released Timestral on October 28, its first time-series foundation models (pretrained models that forecast from raw history without per-dataset training), in four sizes from 5 million to 209 million parameters, with open weights under a modified MIT license. Mistral says the 209M model ranks first among zero-shot models on GIFT-Eval, TIME and fev-bench. One report puts a day-ahead forecast from three weeks of 15-minute data at 140 milliseconds on CPU. The technical report has not shipped, so the rankings are self-reported.

AnalysisCoding interviews stopped telling Felix Rieseberg anything useful this month, so he stopped giving them. Rieseberg, who has led engineering on Claude.ai, Cowork and Claude Code Desktop at Anthropic, wrote on October 27 that the tests were always a weak predictor and that Claude now beats candidates at them. His teams look for small prototypes that show taste and judgment instead, and he admits he has no full replacement yet. His advice to junior engineers is to build a working prototype over a weekend. Claude Code itself started that way, built by two fairly new engineers and shipped about eight days later.

AnalysisA card dispute that used to take up to 90 days now takes about eight at the banks using Casap, the company says, and 73% of those cases never reach a person. Casap announced a $56 million Series B on October 28, led by Base10 Partners, with Experian and Synchrony writing strategic checks. Its agents investigate and decide disputes, fraud claims and chargebacks for more than 100 banks, credit unions and payment firms, and it claims a 96% chargeback win rate against an industry norm near 60%. Collections and insurance claims come next. Every one of those figures is Casap's own.

AnalysisGovernment agencies now pay 40% less for Claude than everyone else. Anthropic expanded its GSA agreement on October 27 to cover federal agencies plus state, local, tribal and territorial governments, applying the discount to Claude Enterprise, Claude for Government and the API, with no minimum spend. Purchases through cloud marketplaces get the same cut in the coming months. Anthropic puts the savings at about $1 billion and says roughly 500 government bodies already use Claude. It lands the day after the company's IPO filing, and a long public-sector customer list reads well in a prospectus.

AnalysisGoogle Cloud revenue grew 103% to $30.7 billion in the September quarter, and Alphabet said sales of TPU systems, its in-house AI chips, were a meaningful part of that. Both giants reported on October 28. Alphabet's first-party model APIs now process nearly 32 billion tokens a minute, four times a year ago, and the Gemini app passed one billion users. Microsoft's Azure grew 48% to $33.2 billion. Spending on property and equipment hit $51.4 billion at Alphabet and $39.4 billion at Microsoft, each roughly double the year before, while total revenue grew 25% and 19%. The build is outrunning the business that pays for it.

AnalysisFor 14 days, very long prompts to one model on Cloudflare's Workers AI could come back carrying fragments of someone else's request. Cloudflare disclosed on October 26 that 155 responses from GLM-5.3 between August 27 and September 9 were affected. Two bugs stacked: a cache pool sized at 681,280 tokens for a model that accepts over a million, and a flaw in SGLang (an open-source inference engine) that passed rejected prompts along instead of returning an error, so the model read stale memory from earlier requests. A researcher found it through Cloudflare's bug bounty, and the fixes went upstream to SGLang.

AnalysisAn AI agent's database connector turned out to work as an open file share. CVE-2026-93328, published October 26 with a 9.3 severity score, hits versions 1.2.0 through 1.4.0 of Google's MCP Toolbox for Databases, a server that lets AI agents query databases over MCP (the Model Context Protocol, a standard for plugging tools into AI models). Path traversal and a missing authorization check in its cloud-storage source let an unauthenticated attacker copy host files to an outside Cloud Storage bucket, or overwrite local files, using crafted tool calls. Version 1.5.0 carries the fix.

AnalysisOne founder, 11 contract labs and under 18 months took a muscular dystrophy treatment from idea to promising mouse results. Anthropic bought that company, Mithril Medicine, and published early results from six drug partnerships on October 27. Founder Cody Tranbarger used Claude to manage contracts, logistics and experiments across the labs. At BigHat Biosciences, Claude-assisted antibody variants bound more than 10 times more tightly than the originals; at Enveda, document review time fell 89%, from over 25 hours to under three. Nearly all of it is preclinical, and the earliest human trial Anthropic names is late 2028.