Cortech Daily

Anthropic strikes a half-price Claude deal with California, a new open model raises the bar for agentic coding, and a wave of vulnerabilities hits AI dev tools, file sharing, and Apple's supply chain.

(0:00) - Intro

(0:21) - Anthropic and California Strike a Half-Price Claude Deal - source

(1:06) - Ornith 1.0: A New Open Model for Agentic Coding - source

(1:58) - Base44 Builds Its Own Model to Stay Defensible - source

(2:47) - Why 'China Matched Anthropic' Doesn't Hold Up - source

(3:39) - AI Agents Are Not Your Coworkers - source

(4:30) - Amazon Q Extension Flaw Exposes Cloud Credentials - source

(5:17) - SimpleHelp Bug Spreads the Djinn Stealer - source

(6:08) - AirDrop and Quick Share Flaws Hit 5 Billion Devices - source

(6:54) - Tata Breach Leaks iPhone 18 Pro Supplier Data - source

(7:44) - Supreme Court Reins In Geofence Warrants - source

(8:34) - Sign-off

Mozilla shows AI coding agents can be tricked into running malware from a clean GitHub repo, fetchmail and n8n scramble to patch nasty bugs, a self-erasing backdoor feeds the ransomware economy, and China tops the supercomputer list without a single GPU.

(0:00) - Intro

(0:23) - Clean repos, hijacked coding agents - source

(1:16) - Fetchmail patches an NTLM overflow - source

(2:08) - n8n's SSRF fix ships off by default - source

(2:57) - A self-erasing backdoor for hire - source

(3:49) - Routing prompts without asking a model - source

(4:39) - Apple's Vision Pro chief heads to OpenAI - source

(5:25) - Apple lobbies for blacklisted Chinese memory - source

(6:08) - China tops the supercomputer list, no GPUs - source

(7:01) - Sign-off

The U.S. government decides who gets to use frontier AI models, fresh research probes whether AI can attack and defend on its own, and a blunt take on what's left of coding as a job.

(0:00) - Intro

(0:24) - Anthropic's Mythos cleared for limited US release - source

(1:13) - OpenAI limits GPT-5.6 and pushes back on the precedent - source

(2:00) - How far behind are open-weight models, really - source

(2:42) - OpenAI, Apple, and SpaceX build their own chips - source

(3:25) - Two thousand people tried to hack an AI assistant - source

(4:10) - Confidence in autonomous AI pentesting cools - source

(4:52) - A targeted developer attack, caught by Claude - source

(5:37) - Vim patches a code-execution flaw - source

(6:19) - Russia tied to the $2.5B Jaguar Land Rover hack - source

(7:07) - Geoffrey Huntley on the end of coding as a job - source

(7:50) - Sign-off

Anthropic accuses Alibaba of the largest distillation attack ever aimed at Claude, the White House slow-rolls the next OpenAI model, new rules arrive for who AI agents are and what they can do, and a fresh round of breaches and patches lands across the security world.

(0:00) - Intro

(0:24) - Anthropic accuses Alibaba of the largest distillation attack on Claude - source

(1:17) - White House asks OpenAI to slow-roll its next model - source

(2:08) - A German ruling puts AI liability on the deployer - source

(2:51) - Paying customers are increasingly picking Claude - source

(3:39) - Patronus raises fifty million to stress-test AI agents - source

(4:32) - A new protocol asks who an AI agent really is - source

(5:23) - The Linux Foundation launches Akrites for open-source security - source

(6:18) - Synology ships a critical fix for MailPlus Server - source

(7:06) - Polymarket refunds users after a third-party breach - source

(7:49) - A former Databricks chief bets on a thousand-fold power cut - source

(8:38) - Sign-off

Anthropic accuses Alibaba of a record Claude-extraction campaign, OpenAI and IBM push new chips, fresh curl and AI-marketplace security findings, and what the latest data says about AI and engineering jobs.

(0:00) - Intro

(0:23) - Anthropic accuses Alibaba of mass Claude extraction - source

(1:14) - OpenAI and Broadcom unveil first custom inference chip - source

(2:05) - IBM claims first sub-one-nanometer chip - source

(2:55) - Aisle finds six new curl flaws, one over 25 years old - source

(3:48) - Unit 42 finds malicious skills in an AI agent marketplace - source

(4:40) - Cloudflare opens self-managed OAuth to all developers - source

(5:33) - Russia used Cellebrite tools after the supposed cutoff - source

(6:26) - New data: engineering is the most AI-resilient job - source

(7:13) - The tokenpocalypse: companies rein in AI spending - source

(7:56) - Simon Willison reshapes a web dataset with Claude Code - source

(8:38) - Sign-off

Anthropic's always-on Claude lands in Slack, a Qwen world-model paper for agents, and a heavy security slate: the Cordyceps supply-chain class, a Five Eyes AI warning, libssh2 CVEs, actively exploited UniFi bugs, the LastPass-Klue breach, and a tightened post-quantum crypto deadline.

(0:00) - Intro

(0:20) - Anthropic puts an always-on Claude inside Slack - source

(1:07) - Qwen team proposes language world models for agents - source

(1:56) - Dries Buytaert rethinks Drupal's place in agentic AI - source

(2:51) - Datasette 1.0 alpha 35 brings schema editing to the browser - source

(3:33) - 'Cordyceps' turns malicious pull requests into supply-chain attacks - source

(4:27) - Five Eyes warns AI hacking is months away, not years - source

(5:16) - Three vulnerabilities disclosed in the libssh2 SSH library - source

(6:10) - CISA flags four actively exploited Lantronix and Ubiquiti bugs - source

(6:57) - LastPass customer data exposed via Klue supply-chain breach - source

(7:43) - White House moves up the post-quantum crypto deadline - source

(8:36) - Sign-off

Anthropic squares off with Washington over a coding model, a fourteen-hundred-dollar experiment out-audits OpenAI's Codex Security, fresh breaches hit Klue and Transport for London, and we go inside the four-hundred-million-dollar machine that prints the world's most advanced chips.

(0:00) - Intro

(0:19) - Anthropic and Washington collide over a coding model - source

(1:13) - Claude may ask to verify your age or identity - source

(2:02) - Simon Willison ports an inpainting model into the browser with Claude Code - source

(2:55) - A new paper reframes prompt injection as role confusion - source

(3:49) - GLM-5.2 takes the crown as the best open model - source

(4:43) - A fourteen-hundred-dollar experiment beats OpenAI's Codex Security - source

(5:40) - OpenAI wants AI to fix vulnerabilities, not just find them - source

(6:34) - A breach at market-research firm Klue ripples into security companies - source

(7:24) - Two Scattered Spider members plead guilty over the Transport for London attack - source

(8:18) - Inside the four-hundred-million-dollar machine that prints advanced chips - source

(9:12) - Sign-off

John Jumper jumps to Anthropic, AutoJack hijacks AI agents, North Korea poisons the Mastra packages, and Secure Boot's clock runs out - your June 21 rundown.

(0:00) - Intro

(0:21) - Nobel laureate John Jumper leaves DeepMind for Anthropic - source

(1:08) - The Atlantic maps the music used to train AI - source

(2:00) - Signal's Meredith Whittaker: AI chatbots are not your friends - source

(2:44) - A field guide to building reliable agentic AI - source

(3:31) - AutoJack: a web page that can hijack an AI agent - source

(4:19) - Microsoft ties Mastra AI package attack to North Korea - source

(5:08) - Klue OAuth breach feeds Salesforce data theft - source

(5:56) - Security week: FortiBleed leak and Splunk under attack - source

(6:45) - ShinyHunters leak exposes Madison Square Garden data - source

(7:30) - Secure Boot certificates start expiring June 24 - source

(8:17) - Sign-off

Anthropic's models become a geopolitical bargaining chip, a Nobel laureate switches labs, and a rough week for web servers, package repos, and satellites.

(0:00) - Intro

(0:22) - Anthropic's Fable 5 hits a wall three days after launch - source

(1:13) - Trump reverses on calling Anthropic a national security threat - source

(1:58) - Nobel laureate John Jumper leaves DeepMind for Anthropic - source

(2:43) - The real value of MCP is isolating auth, not tools - source

(3:32) - The VLC creator is building real-time control for robots - source

(4:21) - Norway draws a hard line on AI in elementary school - source

(5:14) - F5 patches two critical NGINX flaws in HTTP/3 - source

(6:00) - JCPenney breach exposes 368,000 employee records - source

(6:52) - A sprawling supply-chain attack hits the Arch User Repository - source

(7:43) - A satellite reveals the immense scale of GPS tampering - source

(8:39) - Sign-off

A critical Splunk flaw under active attack, a two-billion-dollar bet on world models, a Texas breach exposing three million IDs, a major takedown of the SocGholish malware network, and Interpol's forty-billion-dollar measure of Asian scam compounds.

(0:00) - Intro

(0:21) - Critical Splunk flaw is under active attack - source

(1:11) - General Intuition raises 300 million for world models - source

(1:56) - Texas breach exposes three million IDs - source

(2:41) - Simon Willison launches Datasette Apps - source

(3:31) - Operation Endgame dismantles SocGholish - source

(4:18) - Defense and AI agents headline YC Demo Day - source

(5:02) - Accenture makes a 4 billion dollar OT security push - source

(5:48) - Interpol pegs Asian scam networks at 40 billion a year - source

(6:37) - Containerd patches five flaws beneath Kubernetes - source

(7:19) - Sign-off