the week in AI, briefly. then briefly again.
Artificial Intelligence — briefly, then briefly again · tbb.ceo
listen tothe daily 0:00 –:––
/daily ·11 NOV 2026 ·WEDNESDAY ·2 MIN READ ·5 STORIES

Wednesday, 11 November

DeepMind formally verifies software from first principles; Anthropic expands into collaborative workspace; and the policy apparatus for AI security starts standardising.

01 / The Day

WEDNESDAY 11 NOV 2026, ranked

05

DeepMind's AlphaProof Code verifies Signal crypto at machine level

Google DeepMind released AlphaProof Code, an agentic framework that used multiple Gemini-Carbon-powered prover agents to mathematically verify the security guarantees of the Signal protocol's reference implementation down to raw AArch64 machine code — generating 6.5 million lines of Lean proofs in four weeks, a task that previously took human verification teams years.

  • Prover agents flagged only ~1% of the code for human review; the rest was proved automatically
  • Verification targeted libsignal at compiled binary level, not just source code
  • Extends formal methods to cryptographic security proofs at a scale not previously demonstrated
Why it mattersAutomated machine-level formal verification, if reproducible at scale, changes the assurance ceiling for critical software — a genuine role for AI agents beyond chat.

Anthropic ships Claude Sheets, Whiteboards, and third-party connectors

Anthropic launched Claude Sheets (a spreadsheet canvas with formula support and .xlsx export), Claude Whiteboards in beta (freeform multi-user visual boards), and graduated Claude Data to general availability — alongside interactive third-party connectors that run Atlassian, Adobe, and HubSpot tools side-by-side with Claude in the same interface.

  • Claude Sheets supports formulas, tabs, drop-downs, and .csv/.xlsx export
  • Third-party connectors run external apps in-pane — a direct challenge to Microsoft 365 Copilot
  • Claude Data (formerly Dashboards) now generally available
Why it mattersAnthropic moving from chat into collaborative workspace puts it in direct competition with productivity suites for the first time — a significant strategic expansion beyond model-as-API.

Stanford HAI: scaling bottlenecks are real, inference-time reasoning is the pivot

Stanford's HAI released the quinquennial AI100 report, titled 'The World Has Changed,' finding that the brute-force compute-and-data pretraining scaling paradigm faces structural bottlenecks — specifically data exhaustion and prohibitive training economics — and documenting the industry's shift toward inference-time reasoning architectures, algorithmic density, and multi-agent systems.

  • Panel identified high-quality data exhaustion and training economics as the core scaling ceiling
  • Report charts the pivot to inference-time compute: reasoning models, algorithm efficiency, agents
  • Warns of a widening gap between model deployment velocity and regulatory oversight
Why it mattersAn academic framework corroborating what practitioners have been saying privately becomes public scaffolding — expect this report to shape funding and policy arguments for years.

Adobe and Google DeepMind co-develop next-gen image models for Photoshop

At Adobe MAX, Adobe and Google announced a deep reciprocal enterprise integration: Adobe tools are embedding natively into Google Workspace and Search AI Mode, while Gemini image and video models are being integrated into Photoshop, Express, and Premiere — with an explicit R&D partnership to co-develop next-generation image models for near-real-time prompt-driven editing.

  • Adobe Acrobat and creative connectors land natively in Google Drive, Gmail, and Search AI Mode
  • Gemini image and video models go into Photoshop, Express, and Premiere
  • Joint model training with DeepMind targets real-time prompt-driven edits within Photoshop
Why it mattersThe alliance effectively bundles the two largest non-Microsoft AI productivity suites against Microsoft 365 Copilot — and signals that frontier-model IP is now a currency in enterprise software partnerships.

Frontier Model Forum releases CVSS-style jailbreak rating standard

The Frontier Model Forum — the Anthropic/OpenAI/Google/Microsoft consortium — published Version 1.0 of its Jailbreak Rating Framework for Frontier Cyber Threats, a CVSS-style scoring system that standardises vulnerability disclosure for frontier LLMs across four vectors: capability gain, capability scope, deployment friction, and discovery difficulty.

  • Four scoring vectors give reproducible severity metrics for LLM exploits — borrowing CVSS structure
  • Specifically targets exploits that unlock offensive cyber capabilities, not generic jailbreaks
  • Signals AI security moving from ad-hoc disclosure to standardised, comparable scoring
Why it mattersShared scoring standards change how security researchers prioritise and disclose model vulnerabilities — the same shift that made CVSS foundational to software security now applied to AI.
Be subscriber #028 weekly AI email interest list
Past days

The Daily Archive