03 · The Daily
The Daily · Thursday 12 November 2026
Microsoft researchers described Lazarus, a method for inserting a hidden backdoor into an AI agent by editing fewer than 10 model neurons, without training the model. Microsoft said it has not observed threat actors using the technique. The United States and China expect to exchange advanced AI testing practices and are establishing a channel to alert each other to serious incidents, Treasury Secretary Scott Bessent said.

The day in brief
Microsoft researchers described Lazarus, a method for inserting a hidden backdoor into an AI agent by editing fewer than 10 model neurons, without training the model 1. Microsoft said it has not observed threat actors using the technique 1. The United States and China expect to exchange advanced AI testing practices and are establishing a channel to alert each other to serious incidents, Treasury Secretary Scott Bessent said 2.
Railway estimated stolen customer keys led to about $740,000 in unauthorized spending after an October breach 3. Claroc said 35,578 live apps built with Grok opened without sign-in 5. Google DeepMind researchers used AI agents and human-reviewed specifications to verify security properties of compiled Signal protocol code for one-to-one sessions 4.
Google's Gemini Decisions API evaluates up to 1,000 independent questions about shared multimodal input, returning probabilities and confidence scores instead of answer text 7. Mercor raised $750 million in a Series D at a $20 billion valuation 9. Rox raised $100 million in Series B financing at a $1.3 billion valuation to expand its work on AI agents for enterprise sales teams 6.
The five
- Microsoft researchers show neuron edits can backdoor AI agents
Microsoft researchers describe a method for inserting a hidden backdoor into an AI agent by editing fewer than 10 model neurons. The company says it has not observed threat actors using the technique.
Source · Microsoft Research
- Bessent outlines U.S.-China AI risk talks and planned alert channel
Treasury Secretary Scott Bessent said the United States and China expect to exchange practices for testing advanced AI models. The countries are also establishing a channel to alert each other to serious incidents involving their models.
Source · U.S. Department of the Treasury
- Railway estimates stolen keys led to about $740,000 in unauthorized charges
Railway estimates that stolen customer keys were used to incur about $740,000 in unauthorized spending after an October security breach. The company says variables in at most 1.85% of workspaces could have been decrypted.
Source · Railway
- DeepMind uses AI to verify key properties of Signal protocol code
Google DeepMind researchers used AI agents and human-reviewed specifications to verify security properties of compiled Signal protocol code. The proofs cover one-to-one sessions but have limits, including on claims about quantum attackers.
Source · Google DeepMind
- Claroc finds patient data and passwords in open Grok-built apps
Claroc says 35,578 live apps built with Grok opened without a sign-in, including some containing patient information and employee passwords. SpaceXSI removed 24 apps built by its own and X staff after the researchers reported them, Claroc says.
Source · Claroc
Also today
- Rox raises $100 million to expand enterprise revenue agents
- Google releases Gemini Decisions API for multimodal choices
- Anyday opens free AI exploit catalog, with downloads for verified users
- Mercor raises $750 million to expand enterprise AI evaluations
- Anthropic adds beta sheets and whiteboards to paid Claude plans
- Alibaba's QwenCloud opens Qwen4-Max-Preview through November 25
- Oura agrees to buy AI primary care company Counsel for $625 million
- Torus raises $500 million and launches modular AI compute system
- Google DeepMind releases AlphaGenome 2 for cell-specific research
- UK backs Fractile in AI hardware round worth over £500 million
- Parents sue Google over Gemini AI Mode chats on daughter’s school Chromebook
- Bilt selects Cerebras to power Neighborhood OS and AI concierge
- AWS Adds Usage-Based Billing for SageMaker Streaming Models
- Pew maps data-center growth as surveys show local opposition
- PuppyGraph raises $10 million for enterprise AI agent data tools
- JetBrains adds Claude Agent and Codex to ReSharper
- Study finds swearing in a minority of AI models' benchmark runs
- IBM survey finds gap between confidence and proof of digital control
- Nihon Knowledge launches local LLM tool for software development
- Tether introduces BotPsy, a 0.8B on-device agent model
- TIER IV releases driving AI models for different sensor setups
- Gigabyte says 64GB AI TOP ATOM desktop is now available
- SK hynix earns LPDDR5X safety certification and a cybersecurity certification
- NVIDIA adds AI workload inventory to GPU platform runtime
- Meta says compute allocation added about 10% to ad ranking on top of model and system advances
- Amazon says autonomous AI agents wrote its AWS Emulator
- AWS introduces cross-framework evaluation system for AI agents
- Plurai raises $13 million and opens agent simulation platform
- ElevenLabs releases Scribe v3 transcription models, citing fewer errors
- Hugging Face details pooling-aware training for embedding models
- Parloa and OpenAI Expand Customer-Service AI Partnership
- Sakana AI names former NSA chief and CIA officer as advisors
- Databricks previews publishing metric views to Power BI
- Hitachi Vantara adds four-node starting option for AI storage
- Embed Security lets customers create AI investigation agents
- Lava Security's neocloud and AI data center risk list joins OWASP as an incubator project
- rakumo brings AI agent to invoice and estimate approval workflows
- Ampt puts 80 kW battery converter for AI data centers into pilots
- Cursor rolls out shared Projects for teams in beta
- Apple’s Claris previews Harbor for running AI models on Macs
- Altius Link starts AI voice service for contact centers
- Google adds offline speech input and playback to AI Edge Gallery
- Armada and Johnson Controls start Dublin data center production
- SolidRun launches i.MX 95 platform with Holoscan demonstration
- DeepL launches Aure voice translation model
- Rackspace and Together AI plan dedicated GB300 inference deployment
- Ulys opens prompt-based investing app to the public
- Vercel adds MiniMax H3 Image models to AI Gateway
- JetBrains MPS 2026.2 adds tools for coding agents
- Ancep raises $8.5 million for insurance-broker AI agents
- Sandia reports about 90%-plus accuracy in continual-learning test
- GoodRx connects prescription prices to Gemini for healthcare preview
- Workato connects with SoftBank’s AGENTIC STAR platform
The paper of the day
On X today
- Jason Wei (@_jasonwei)
Created a "benchmark benchmark": a frequency distribution of which benchmarks were reported by the frontier labs in 2026: claude.ai/artifact/22cBP… Some observations: - HLE, SWE-Bench Pro, DeepSWE, and Terminal-Bench saw universal adoption - Scale AI is behind two of those four
The event of the day
Every issue
- Sunday 15 November 2026 — ChatGPT Wallet lets US Pro subscribers approve agent purchases
- Saturday 14 November 2026 — Fed's November survey ranks AI first among financial-stability risks
- Friday 13 November 2026 — Epoch AI says test agents attempted exploits on its live website
- Thursday 12 November 2026 — Nvidia CEO Jensen Huang resigns; Colette Kress takes interim role
- Wednesday 11 November 2026 — Microsoft researchers show neuron edits can backdoor AI agents
- Tuesday 10 November 2026 — TSMC board approves up to $500 million each for Anthropic and OpenAI stakes
- Monday 9 November 2026 — Malbear Labs reports two-week DeepSeek telecom intrusion
- Saturday 7 November 2026 — PSI research system proves Aaronson-Ambainis conjecture
- Friday 6 November 2026 — DOJ Launches Super Intelligence Task Force
- Thursday 5 November 2026 — Acronis Threat Research Unit Identifies Single Operator Using BOSSAI, a 12-Agent AI Mesh
- Wednesday 4 November 2026 — Anthropic Says Claude Leads 26% of AI Research, None Autonomously
- Tuesday 3 November 2026 — EU Court Adviser Says Gemini Answers May Breach Publisher Rights
- Monday 2 November 2026 — Marvell Agrees to Buy Lumentum in $115B All-Stock Deal
- Sunday 1 November 2026 — Chainalysis Discloses Breach Exposing Customer Names and Emails
- Saturday 31 October 2026 — Microsoft Finds North Korean Groups Using ARTEX AI Agents
- Friday 30 October 2026 — Accenture Contractor’s Claude Agent Ran on a Pentagon Network for Two Months
- Thursday 29 October 2026 — Cohere targets C$7 billion in sovereign AI funding round
- Wednesday 28 October 2026 — Google Launches Gemini 4 Carbon With Cybersecurity Focus
- Tuesday 27 October 2026 — Australia’s ASD Says AI Is Speeding Up Cyberattacks
- Monday 26 October 2026 — South Korea Refers Detained AI-Theft Ringleader to Prosecutors
- Sunday 25 October 2026 — GPT and Grok Pick Opposite Parties in 2026 Election Test
- Saturday 24 October 2026 — Texas Seeks to Bar Minors From ChatGPT and Review New Models
- Friday 23 October 2026 — AI-Assisted Proof Resolves 31-Year-Old KLS Conjecture
- Thursday 22 October 2026 — AI-Assisted Proof Resolves 31-Year-Old KLS Conjecture
- Wednesday 21 October 2026 — Radical Numerics Gives AI-Designed Phage Cocktail to a Patient
- Tuesday 20 October 2026 — Ofcom Weighed Urgent Court Action Over Grok Deepfakes
- Monday 19 October 2026 — Judge Keeps Former OpenAI Engineer in Custody in Saudi-Linked Case
- Sunday 18 October 2026 — Tencent Says Agents That Probed Amap Came From Third-Party Model Evaluations
- Saturday 17 October 2026 — Bitsight Finds MONST3R-AI Tool Uses Stolen Keys for Scans
- Friday 16 October 2026 — China Requires AI Export Registration as It Drafts New Curbs
- Thursday 15 October 2026 — FTC Seeks Safety Records From OpenAI and Anthropic
- Wednesday 14 October 2026 — Eight States Sue OpenAI Over ChatGPT Safety
- Tuesday 13 October 2026 — MI5 Launches AI Threat Assessment Centre With Three Partner Bodies
- Monday 12 October 2026 — Anthropic Finds Small Models Near AI-Worm Self-Replication
- Sunday 11 October 2026 — Anthropic Cuts Live Internet Access to Internal Claude Evaluations
- Saturday 10 October 2026 — Anthropic Cuts Live Internet Access to All Internal Evaluations
- Friday 9 October 2026 — Senators Say Data Centers Shift Grid Costs to Households
- Thursday 8 October 2026 — Zenity Says One Prompt Reached Every Agent in an AWS Account
- Wednesday 7 October 2026 — OpenAI Brings GPT-6 Interactive Interfaces to ChatGPT
- Tuesday 6 October 2026 — Mistral Unveils Large 4, a 1.05 Trillion-Parameter Model
- Monday 5 October 2026 — IWF Finds 6,310 AI-Generated Child-Abuse Images in Six Months
- Sunday 4 October 2026 — White House Gives Jay Clayton 120 Days to Assess AI Risks
- Saturday 3 October 2026 — Bessent Discusses AI Incident Channel After U.S.-China Summit Deal
- Friday 2 October 2026 — Nebius Buys Inferize to Reduce AI Model Cold Starts
- Thursday 1 October 2026 — FTC Opens AI Safety Probe Into OpenAI and Anthropic
- Wednesday 30 September 2026 — Trump and AI CEOs Sign Voluntary Safety Accord
- Tuesday 29 September 2026 — UK Tests Put GPT-6 Astra’s Simulated Attack Rate at 29.2%