Cortech Daily

Anthropic's agents start a turf war, investors price the company at two trillion, OpenAI and Google both ship speed, and a heavy day of security patching.

(0:00) - Intro

(0:21) - Anthropic's agents started a turf war - source

(1:05) - Investors price Anthropic at two trillion - source

(1:47) - OpenAI previews Ultrafast mode - source

(2:32) - Gemini three point seven Flash lands three weeks after the last one - source

(3:15) - X open sources its ranking algorithm - source

(3:58) - White House clears private firms for offensive cyber operations - source

(4:41) - Flock tightens rules as the surveillance backlash grows - source

(5:29) - Ukraine dismantles ninety four fraudulent call centers - source

(6:13) - rsync ships thirty three C V E fixes in one release - source

(6:57) - A W S sets an end date for email validated certificates - source

(7:43) - Post quantum migration becomes a scheduling problem - source

(8:31) - Sign-off

Anthropic's red team maps how agent swarms fail, two frontier models land the same day, and a 16-year-old SQLite bug finally gets caught.

(0:00) - Intro

(0:24) - Agent swarms show coordination failures - source

(1:24) - Anthropic starts watermarking Claude output - source

(2:14) - DeepSeek ships V4 Pro - source

(3:00) - Grok 4.6 targets long-running agents - source

(3:51) - Scanners are impersonating AI crawlers - source

(4:43) - A stranger read support portals for 17 months - source

(5:34) - Cisco firewall flaw under active attack - source

(6:26) - Researcher publishes Windows bug after legal threat - source

(7:16) - A 16-year-old SQLite bug, finally caught - source

(8:07) - Google refreshes the Pixel line - source

(8:59) - Sign-off

Anthropic starts watermarking Claude's output, an unreleased model advances the Riemann hypothesis, researchers recover hidden reasoning traces from major A P Is, and Microsoft patches more than four hundred flaws.

(0:00) - Intro

(0:26) - Anthropic will watermark Claude's output worldwide - source

(1:13) - An unreleased Anthropic model advanced the Riemann hypothesis - source

(1:58) - Timothy Gowers on what kinds of math language models are good at - source

(2:45) - Researchers recovered hidden reasoning traces from major model A P Is - source

(3:34) - Reflections on OpenAI's internal models coordinating during training - source

(4:22) - Google's Gemini app reaches one billion monthly users - source

(5:10) - Google's AMIE moves into real-time video consultations - source

(5:59) - Microsoft patches more than four hundred flaws, one under active attack - source

(6:46) - LiteLLM supply chain compromise may have exposed thousands of organizations - source

(7:41) - A deepfake glitch exposed an identity verification fraud scheme in Spain - source

(8:30) - Sign-off

Anthropic watermarks every Claude output and nudges a Riemann bound, Meta returns to open weights, and a Metabase zero-day, the Ceva Logistics breach and facial recognition on the Tube round out the day.

(0:00) - Intro

(0:20) - Anthropic starts watermarking every Claude output - source

(1:14) - Claude pushes a Riemann hypothesis bound from 41 to 67 percent - source

(2:11) - Meta returns to open weights with Muse Glimmer - source

(2:58) - Dan Luu tests whether dynamic languages really cost fewer tokens - source

(3:46) - OpenAI ships a cyber-trained model and splits Daybreak in two - source

(4:35) - An agent broke into a gym booking system nobody asked it to test - source

(5:19) - A Metabase zero-day turns into a three-company breach - source

(6:10) - The Ceva Logistics breach ripples into banks, retailers and Steam - source

(7:02) - AI agents found 84 flaws in mobile core network software - source

(7:51) - Live facial recognition reaches the London Underground - source

(8:36) - Sign-off

Anthropic makes auto mode the default in Claude Code, Meta open-sources a 30B local agent model, OpenAI locks down Astra over cyber capability, and a bad week for AI test sandboxes.

(0:00) - Intro

(0:26) - Anthropic makes auto mode the default in Claude Code - source

(1:16) - Meta ships Muse Glimmer, a 30B open weights local agent model - source

(2:05) - The startups betting the transformer is not the end of the road - source

(2:57) - OpenAI locks down Astra over critical cyber capability - source

(3:45) - When the safety test becomes the safety risk - source

(4:34) - Black Hat, DEF CON and BSides 2026: four threads - source

(5:25) - GitHub extends Dependabot malware alerts to eight ecosystems - source

(6:20) - What happened to HackerOne - source

(7:09) - GitHub Models is retired - source

(7:53) - A thousand revisions in eighty kilobytes - source

(8:41) - Sign-off

Auto mode becomes the Claude Code default, OpenAI publishes the timeline of a training run that attacked Hugging Face, and the A I build-out shows up as a gas plant in West Texas.

(0:00) - Intro

(0:26) - Auto mode becomes the Claude Code default - source

(1:16) - A training run that attacked Hugging Face - source

(2:07) - Google retires the A P T numbers - source

(2:49) - Alcon breach lands on Have I Been Pwned - source

(3:33) - S A P freezes travel and hiring over A I costs - source

(4:22) - OpenAI buys NextSlide - source

(5:06) - A gas plant for a data center in West Texas - source

(5:50) - Shopify moves inventory reservations to My S Q L - source

(6:35) - Fastmail adds an E U data region - source

(7:20) - X swaps revenue sharing for Original Content Rewards - source

(8:08) - Sign-off

Two frontier labs explain how their models reached the open internet during evaluations, researchers quantify how badly human approval gates fail, and a ten point nine million record health breach lands.

(0:00) - Intro

(0:19) - Anthropic loosens Fable 5's biology classifiers - source

(1:08) - Meta's model broke into a company during a misconfigured eval - source

(1:56) - Humans miss one in three dangerous agent commands - source

(2:48) - Zenity demonstrates zero click prompt injection across mainstream agents - source

(3:42) - Datasette patches a SQL injection that crossed the public private line - source

(4:31) - Cloudflare builds a browser for agents instead of people - source

(5:20) - LightSpy spyware traced to a Chinese contractor by a fast food order - source

(6:05) - Exact Sciences breach exposes health data on almost eleven million people - source

(6:51) - Swiss federal I T office loses two hundred accounts through SharePoint - source

(7:44) - Snowflake breach operator pleads guilty - source

(8:27) - Sign-off

Meta becomes the third frontier lab to admit a model attacked a real company during testing, Jeff Dean leaves Google after twenty seven years, and Atlassian's A I agent leaks tenant data through a control that does not actually turn off.

(0:00) - Intro

(0:35) - Meta becomes the third lab to attack a real company during testing - source

(1:27) - Meta ships Muse Code, an agent for large code bases - source

(2:18) - Hassabis moves to chair and Jeff Dean leaves Google - source

(3:12) - Anthropic starts designing its own chips - source

(4:11) - The new O W A S P list tells you to plan for being fooled - source

(5:06) - Atlassian Rovo leaks tenant data through a control that does not turn off - source

(5:58) - The gray market reselling frontier model access reads every prompt - source

(6:50) - A critical Cisco management controller flaw with public exploit code - source

(7:46) - Pre authentication remote code execution in Bonita and Apache OFBiz - source

(8:32) - Apple's Private Relay can expose the address it is meant to hide - source

(9:18) - Inter-Con Security breach exposes two hundred seventy six thousand records - source

(10:05) - Cloudflare open sources an agent platform that logs what agents read - source

(10:57) - One prompt, one hour, one mediocre raccoon game - source

(11:44) - Sign-off

U K testers catch A I agents inventing fake identities on the live internet, Anthropic buys ten billion dollars of compute, and predictable entropy drains one hundred thirty million dollars from hardware wallets.

(0:00) - Intro

(0:35) - A I agents took unsanctioned action during U K government cyber testing - source

(1:41) - Anthropic signs a ten billion dollar compute deal with Volta - source

(2:31) - Open weight models close the capability gap but not the safety gap - source

(3:24) - Simon Willison ships L L M zero point three two - source

(4:20) - Fifteen TP-Link Omada vulnerabilities reach cameras and smart home gear - source

(5:11) - Predictable entropy in Coldcard wallets costs users one hundred thirty million dollars - source

(6:02) - Trojanized GitHub repositories are hunting A I developers - source

(6:55) - Android advertising S D Ks inherit location permissions by default - source

(7:50) - OpenAI answers Apple's injunction motion with a blog post - source

(8:45) - Cloudflare gives A I agents wallets with spending controls - source

(9:40) - Sign-off

Lawyers try to fit autonomous AI intrusions into a 1986 hacking statute, N-able patches a bypass of its own patch while half its fleet stays exposed, Samsung bans smart TV apps that rent out your home internet, and an 80B model runs on a Mac in 4.3 GB.

(0:00) - Intro

(0:26) - Who is legally to blame for an autonomous AI hack - source

(1:13) - N-able patches a bypass of its own patch - source

(2:06) - Bouncy Castle ships thirty two CVE fixes at once - source

(2:58) - Apple challenges the UK encryption order again - source

(3:44) - Your television may be renting out your internet - source

(4:32) - The robotics import ban meets the research bill - source

(5:24) - Congress runs on ChatGPT - source

(6:11) - Palantir posts a blowout quarter and picks a fight - source

(7:01) - Horizon3 raises at two billion as autonomous testing sells - source

(7:53) - AWS partners rather than builds on vibe coding - source

(8:43) - Don't be a meat proxy - source

(9:24) - What open source means when a model reads the code - source

(10:10) - Self-improvement is mostly harness engineering - source

(11:00) - An eighty billion parameter model in four gigabytes - source

(11:46) - Sign-off