AI News

AI news brief: Friday, November 20, 2026

36 stories, most important first.

OfficialSafety & Security · · Updated

MiniMax says its Cyber model found practical breaks in several cryptographic schemes

MiniMax says its Cyber model automatically found flaws and built working attacks on NIST candidates and peer-reviewed cryptographic schemes.

Why it matters: The claims suggest AI models can now find and exploit flaws in cryptographic designs and reference code at speed. That shifts the bottleneck toward human verification of security margins. Teams planning post-quantum migration should expect more cryptanalysis results to arrive faster.

1 source
ReportedResearch · · Updated

Levent Alpöge says he and Claude found a proof of the Riemann Hypothesis

Mathematician Levent Alpöge says he and Claude discovered a proof of the Riemann Hypothesis, and that the method extends to the Generalized Riemann Hypothesis.

Why it matters: If the claim holds, it would show AI systems contributing to major open mathematics. For now it is one researcher's unverified assertion, so companies should not treat it as established until expert mathematicians have checked it.

3 sources
ReportedProducts · · Updated

Google brings computer-use capability to Gemini for Mac

Google says Gemini for Mac can now navigate apps and complete multistep tasks while users work on other things.

Why it matters: Desktop agents that operate apps on a Mac raise new questions about data access and oversight for employees using Gemini on work machines.

2 sources
OfficialProducts ·

Google rolls out AI-focused search homepage and removes the Search button

Google is deploying a redesigned desktop search page with AI chat access and buttons for image generation, file analysis and Gemini brainstorming, and no separate Search button.

Why it matters: Google is moving the starting point of search toward AI features, which could change how users reach websites and how much traffic reaches sites that depend on Google search.

1 source
OfficialChips & Infrastructure ·

Meta reconsiders 2030 net-zero goal as AI data center emissions hit a record

Meta's 2026 sustainability report says it is re-evaluating its 2030 net-zero goal after a record rise in greenhouse gas emissions tied to AI data center growth.

Why it matters: Companies that set emissions targets around AI infrastructure may see those targets slip as compute demand grows. Meta's revision shows how quickly that pressure can arrive. Buyers and partners should expect sustainability commitments from cloud and AI suppliers to be revised.

1 source
ReportedSafety & Security ·

Researcher seeds 2 million canary secrets into AI routers and finds data leaking

HD Moore planted 2 million traceable fake secrets in prompts sent through AI routers and reports that sensitive information leaked widely along the chain.

Why it matters: Companies that send prompts through third-party routers or gateways may expose credentials and private data to intermediaries they do not directly control. The experiment suggests such leaks can occur along the chain of services.

1 source
ReportedBusiness & Funding ·

Anthropic's IPO could eventually steer up to $156B to charity

Reports say founder pledges, employee donations and a company fund could direct as much as $156 billion to charity, about a quarter of US giving in 2025.

Why it matters: Funding for AI safety and other causes may depend on how much newly liquid Anthropic wealth is actually donated.

1 source
ReportedBusiness & Funding ·

Blackstone trust to buy 75.1% stake in 13 Amazon data centers for $5B

Blackstone Digital Infrastructure Trust is acquiring a 75.1% stake in 13 Amazon data centers with 647 MW of capacity for $5 billion.

Why it matters: Long-term leases on AI data centers shift the cost of capacity from flexible cloud rental to fixed commitments, which could affect how AI companies plan their budgets.

1 source
OfficialResearch ·

METR says AI time horizons on R&D tasks double about every 3.6 months

METR's evaluations find frontier models usually match or beat skilled humans on coding and optimization, with time horizons doubling roughly every 3.6 months.

Why it matters: Companies relying on AI for research and engineering work should not assume older benchmarks still bound risk, because METR says capability depends on the specific task profile. Evaluators' trend estimates are also sensitive to method choices, so single-number claims about progress deserve caution.

1 source
ReportedResearch ·

Theorem and Anthropic report LLM-scaled formal verification is getting cheaper

Theorem and Anthropic describe LLM-scaled formal verification and say Opus 5.5 reached a 50% solve rate on small tasks with about 40% fewer tokens than Opus 5.

Why it matters: Formal verification is moving closer to being affordable for security work on real codebases, which could change how teams check AI-written code.

1 source
ReportedResearch ·

Vals says GPT-6.2 Astra produced a Lean proof of Smoktunowicz's Gap Theorem

Vals believes GPT-6.2 Astra produced the first Lean proof of Smoktunowicz's Gap Theorem, a result made this week.

Why it matters: If the claim holds, frontier models are producing machine-checked proofs of research-level results, which strengthens the case for verifying AI output formally.

1 source
ReportedAI in Practice ·

Ramp says its Inspect coding agent now writes about 75% of its code

Ramp engineer Zach Bruggeman says Ramp's internal Inspect coding agent writes and ships about 75% of its code, and that Inspect is now available outside Ramp via a waitlist.

Why it matters: The claim gives a concrete internal benchmark for how much of a production codebase an agent can write, though it comes from the company itself and is unverified. Teams weighing shared coding agents may find the emphasis on public sessions and reviewable verification relevant.

1 source
ReportedResearch ·

Microsoft Research's LongCapHarness boosts long-context scores without training

Microsoft Research's LongCapHarness changes the code around a frozen model, splitting tasks into smaller operations, and reports large gains on long-context tests.

Why it matters: Teams working on long-context retrieval may be able to fix some failures through workflow design before committing to fine-tuning. The gains depend on the model and task.

1 source
OfficialAI in Practice ·

Stanford researchers unveil TRACER, an AI system that reviews ICU video to flag delirium

A Stanford-led team describes TRACER, which combines continuous ICU bedside video with electronic health records to help clinicians spot early signs of delirium.

Why it matters: Hospitals considering bedside video AI now have a published prototype that links continuous activity data to clinical records. Its 61% query accuracy shows it is far from clinical grade. Its training-free design, which the authors say avoids patient-data training, could make cross-site deployment easier if validation holds up.

1 source
ReportedSafety & Security ·

ChatPRD's Claire Vo says token fraud forced her to end free trials

Claire Vo says abuse of the ChatPRD free trial, often aimed at reselling tokens, has led her to wind down free trials entirely.

Why it matters: Companies offering free AI trials or promo codes can have their compute resold or used for unrelated commercial work, which raises costs for the provider. Prompt patterns may also be used to identify which underlying models are being served.

1 source
OfficialProducts ·

OpenAI's Codex CLI 0.171.0 adds GPT-6.2 Sol and Astra and a default TUI workspace

Codex CLI 0.171.0 adds GPT-6.2 Sol and Astra to model catalogs, turns on the TUI workspace by default, and adds configurable undo and redo.

Why it matters: Teams using Codex CLI may find a new default model and a different default interface after upgrading. The MCP sign-in fix matters for enterprises whose identity providers issue no refresh tokens.

1 source
OfficialProducts ·

Hermes Agent v0.23.3 ships a stable patch covering about 370 merged pull requests

Nous Research's Hermes Agent v0.23.3 is a stable patch release for the CLI, Microsoft Store app, Docker and Hermes Cloud, bundling about 370 merged PRs since v0.23.2.

Why it matters: Teams running Hermes in Docker or Hermes Cloud get a stable tag. The cache default change, however, could alter cost and latency in long-running agent sessions. The release notes are thin on detail, so teams cannot yet verify the behaviour from them alone.

1 source
ReportedProducts ·

Replit says its agents rebuilt a rival's features 1:1 after studying its workflows

Replit says its agents studied thousands of a competitor's workflows and rebuilt the competitor's features 1:1, while pitching its DataCat product against an $83,700 monitoring bill.

Why it matters: Buyers who rely on AI agents to copy or reproduce competitors' products should ask what access those agents had to third-party systems and data.

1 source
ReportedProducts ·

Hebbia reports better financial research results with prerelease GPT-6 Astra

Hebbia says its prerelease tests of GPT-6 Astra found gains in source coverage, citation relevance and presentation quality for financial research.

Why it matters: Firms that use AI for financial research may find citation accuracy and source coverage the most useful measures of a model. This account reports gains on both, though only from one customer's prerelease tests.

1 source
ReportedModels ·

Thinking Machines releases open-weights Inkling 2 Mini, a small looped transformer

Thinking Machines has released Inkling 2 Mini under Apache 2.0, and Artificial Analysis gives it a score of 20 on its Intelligence Index with 5.4B active parameters.

Why it matters: Teams that want to run a capable model on a laptop or small server get an openly licensed option. Artificial Analysis rates it as cheaper and faster per task than Qwen3.8 27B at a similar score. The figures are Artificial Analysis's own benchmark results.

1 source
OfficialProducts ·

Claude Code 2.1.330 adds sign-in for Enterprise Admin Console organizations

Anthropic's changelog says that after claude auth login, Claude Code applies the organization's settings and uses its cloud provider or gateway.

Why it matters: Companies that manage Claude Code centrally can apply organization settings and route traffic through their own cloud provider or gateway after one sign-in step.

1 source
OfficialSafety & Security ·

Claude Code 2.1.330 asks approval for more root-level rm commands

Anthropic's changelog says more rm commands aimed at the root, home, a top-level folder or the working folder now ask for approval.

Why it matters: Agent workflows that run shell commands get stricter default protection against deleting key directories, which may stop some unattended jobs.

1 source
OfficialProducts ·

Claude Code 2.1.330 adds easel, a built-in chart mod for paid plans

Anthropic's changelog says easel, a built-in mod, draws charts on request or when MCP metrics results arrive, and automatic charts can be turned off in /config.

Why it matters: Developers who run data-heavy sessions in Claude Code now get built-in charting. That can change how results appear in their workflows.

1 source