AI Daily Brief · Nov 7, 2026 | BioBasicsBench; OpenCode Browse; DeepMind RSA cryptanalysis; Meta Muse specialized models; PSI Lean proof claim

Over the past 24 hours, benchmarks and agent tooling landed together. OpenAI released BioBasicsBench for routine wet-lab analysis and joined the Frontier Model Forum’s cyber jailbreak rating framework. Open-source coding-agent maker OpenCode shipped Browse—an MCP server plus browser extension that reuses your real logins and hands sensitive steps back to you. Google DeepMind said human-supervised AI agents helped uncover weaknesses tied to legacy 1024-bit RSA, while stressing that modern 2048-bit and larger deployments are unaffected. Meta rolled specialized models trained with Muse Wonder into Muse and Meta AI. Physical Superintelligence claimed an internal system running a non-public Anthropic model produced a Lean-accepted proof of the Aaronson–Ambainis conjecture. On the agent side, Claude Code cloud sessions gained easier environment setup, Apiiro reported credentials leaking via settings files on public GitHub, and Amp, OpenRouter, and GitHub each shipped agent-scale product or infrastructure updates.

Today’s highlights

• BioBasicsBench: 100 wet-lab tasks; GPT-6 Astra ~66.7% at xhigh [Company claim].
• OpenCode Browse: real local browser + MCP, with handoff for login/2FA [Company claim].
• DeepMind: human-guided agents probe legacy 1024-bit RSA; 2048+ unaffected [Company claim].
• Meta Muse: Active Memory / Active Focus specialized models for Muse and Meta AI [Company claim].
• Claude Code: cloud sessions auto-configure deps and API keys; Apiiro reports secrets in public settings files [Company claim / security-vendor report].

1. OpenAI launches BioBasicsBench for everyday wet-lab analysis; also backs FMF jailbreak rating framework [Company claim]

Benchmark: OpenAI, with enzyme-design company Ligo, introduced BioBasicsBench—100 tasks across protein methods, imaging and cytometry, quantification and assays, nucleic-acid methods, and molecular cloning. Agents must read raw gels, plates, microscopy, and instrument outputs, spot errors, and refuse unsupported conclusions. At the xhigh reasoning setting, OpenAI reports GPT-6 Astra earned a mean 66.7% of available credit (95% CI about 60.3–73.1%), versus 43.8% for GPT-5.6 Sol, across seven models, 37 configurations, and about 18,500 results. The company says models still stumble on controls, imaging, and experimental context.

Also: OpenAI and the Frontier Model Forum released Version 1.0 of the Jailbreak Rating Framework for Frontier Cyber Threats, scoring cyber jailbreaks on capability gain, scope, ease of deployment, and ease of discovery. Public comment runs through Dec 21, 2026 [Company / industry claim].

Sources: @OpenAI 1 (~3,229 likes) · @OpenAI 2 · BioBasicsBench post · FMF framework · Unite.AI

2. OpenCode ships Browse: local browser + MCP with your existing logins [Company claim]

Product: OpenCode says Browse pairs an MCP server with a Chromium extension so agents can use your real browser (Chrome, Edge, Brave, Arc, and others)—tabs, sessions, and extensions included. Agents open background tabs in collapsed groups and drive them with Playwright. Logins, 2FA, and payments hand the tab back until you click Continue. Features include screenshots, recordings, HAR capture, and site allowlists. It works with OpenCode, Claude Code, Cursor, Codex, Zed, and other MCP clients.

Also: A developer noted Browse can run the agent and the browser on different machines.

Sources: @opencode (~3,859 likes) · @thdxr · OpenCode Browse · GitHub anomalyco/browse

3. Google DeepMind: human-guided AI agents surface weaknesses in legacy 1024-bit RSA [Company claim]

Research: DeepMind says models under human supervision helped find weaknesses related to legacy 1024-bit RSA. On the signature side, once an attacker controls roughly half or more of the signed block, padding may add little asymptotic protection; agents reproduced a published benchmark instance with about 3.5× less CPU time. A second working paper discusses “fan factories” for batch factoring; company messaging and secondary coverage often cite roughly 30× lower average cost per additional 1024-bit modulus as an asymptotic/heuristic estimate. Figures such as ~1,600× are largely theoretical, not end-to-end measured runs [Company claim; defer to the working papers]. DeepMind states modern 2048-bit and larger RSA deployments are unaffected by these specific results.

Sources: @GoogleDeepMind · @pushmeet · Signature-side PDF · Batch-factoring PDF · CryptoBriefing

4. Meta brings Muse Wonder-trained specialized models to Muse and Meta AI [Company claim]

Product: Meta says specialized versions of its existing models, trained using Muse Wonder, are rolling out to Muse and Meta AI with Active Memory and Active Focus. The research blog reports about a 2.4× reduction in forgetting information critical to goals in Muse, about a 67% improvement on everyday tasks when constraints change, roughly 90% correct use of critical context on an Active Memory eval (prior ~76%), and a Personal Superintelligence Index score of about 47% for the current Muse model versus about 38% previously [Company claim].

Sources: @AIatMeta · Meta research blog

5. Physical Superintelligence claims a Lean proof of the Aaronson–Ambainis conjecture [Company claim]

Claim: PSI says an internal research system running a non-public Anthropic model produced a complete proof around day three that Lean accepted, then improved the result twice with Lean accepting the finals. The conjecture, open since around 2008, bears on whether quantum algorithms can be exponentially faster than classical ones on random inputs. PSI published a research page plus a Lean 4 formalization on GitHub; the repo says the key theorems have no placeholders or extra axioms. [Company claim] Independent peer review was not found for this brief; the README still leaves statement fidelity to human readers. The company’s own announcement post landed just before this edition’s window; discussion continued over the past 24 hours.

Sources: @kimmonismus · PSI research page · GitHub Lean formalization · Business Insider

6. Claude Code: easier cloud-session environment setup [Company claim]

Feature: Anthropic says Claude Code cloud sessions on Pro, Max, and Ultra can figure out dependencies, network access, and API keys, then save the environment for next time. The company says API keys are stored securely so Claude never sees them in the clear.

Sources: @ClaudeDevs (~6,673 likes) · @testingcatalog · Cloud environments docs · Claude Projects

In brief

• Apiiro: Claude Code settings secrets on public GitHub [Security-vendor report / media]: About 1,122 public .claude/settings.local.json copies held at least one secret; broader scans reached about 2,539 secrets across 1,994 repos. Rotate first, then clean history. (@TheHackersNews · The Hacker News)
• Amp Spaces [Company claim]: A higher-level workspace where one agent spawns and coordinates others. (@lewis_b_metcalf · @thorstenball · Amp: Spaces)
• OpenRouter Auto Router / Auto Code [Company claim]: Auto Code said to approach Opus 5.5 quality at about 74% lower cost, p90 ~0.65 s. (@OpenRouter)
• json-render/chat [Open-source author claim]: Generative UI cards for Slack, Teams, and more. (@ctatedev · GitHub)
• GitHub’s new Git storage architecture [Company claim]: Internal tests cited ~5,000 concurrent agents and ~7,200 pushes/s on one repo. (@github · GitHub Blog)
• Trending agent / Skills repos: security-audit-skill, orca, auto-ontology, agent-browser, google/skills, anthropics/skills; HN Agent Memory ~79 points. (security-audit-skill · orca · auto-ontology · agent-browser · google/skills · anthropics/skills · Agent Memory)
• NVIDIA Lyra 2.0 [Company claim]: Single image to explorable 3D worlds. (@NVIDIAAI · project page)
• Mistral Voxtral 3 end-of-turn [Company claim]: (@MistralAI)
• Anthropic science-workload data [Company / research claim]: Median ~38% of researcher time if widely adopted. (@emollick · Anthropic research)
• GPT-6.2 / Fable “ships today” [Unverified / community rumor]: Watchers still waiting; no matching flagship posts verified for this edition.
• browser_use accessibility-tree tip [Tip]: (@browser_use)
• @theo on the engineer gap [Opinion]: Said AI has widened the gap to at least 100×. (@theo)
• Mastra / LangChain updates [Company claim]: (@mastra · @LangChain)

---

About this column and how we verify

1. AI-assisted editing: AI Daily Brief is AI-assisted. This edition was produced with help from Grok Bot and covers roughly 00:00–22:06 Beijing time on 2026-11-07.
2. Tweet-by-tweet checks: Every cited post was fetched individually to confirm it exists, the author matches, and the timestamp falls in this edition’s window; all selected posts also received a second check via the X API. Like and view counts are readings from about 22:06 Beijing time.
3. X News summaries are leads only: X News blurbs are Grok-generated and used only to find stories. Facts come from original posts, official pages, and media; when a summary conflicts with a primary source, the primary source wins.
4. Label meanings: [Company claim] means figures or product details from the company, not necessarily independently re-measured; [Security-vendor report / media] means a security firm’s scan as relayed by media; [Unverified / community rumor] means community monitoring without a matching official flagship post; [Opinion] means a personal judgment.