Cortech Daily

China's GLM-5.2 tops the open-weight charts, Nvidia's coding agents train robots to 99 percent, GitHub feuds over a supply-chain worm, and Pew finds only 16 percent of Americans optimistic about AI.

(0:00) - Intro

(0:24) - GLM-5.2 tops the open-weight leaderboard - source

(1:13) - Vercel's lesson: delete most of your agent's tools - source

(2:03) - Zvi on the Fable export-control fight - source

(2:52) - Nvidia's coding agents train robots to 99 percent - source

(3:37) - Who watches AI agents writing lab code? - source

(4:28) - Google's standard for agents to find and trust tools - source

(5:20) - FortiBleed: tens of thousands of firewalls swept up - source

(6:13) - GitHub disputes the flaws fueling a supply-chain worm - source

(7:00) - ESET details the GentleKiller EDR-killer - source

(7:49) - Pew: only 16 percent of Americans see AI upside - source

(8:38) - Sign-off

Export-control fights over Anthropic's Fable 5, ChatGPT slipping below half the market, a one-click Copilot data leak, Chinese espionage inside US research labs, and a brain implant that gives a man his voice back.

(0:00) - Intro

(0:22) - Export controls on Fable 5 could backfire on cyber defense - source

(1:10) - ChatGPT slips below half the assistant market - source

(2:05) - Salesforce buys AI support platform Fin for 3.6 billion dollars - source

(2:56) - A new open protocol for how AI agents register users - source

(3:50) - Copilot SearchLeak flaw could turn one click into data theft - source

(4:43) - Cisco discloses a second exploited SD-WAN flaw in two weeks - source

(5:38) - 56 million accounts surface in fresh stealer-log dump - source

(6:30) - China-linked group spied on US research labs for over a year - source

(7:25) - Software supply chains face a transparency test - source

(8:20) - An ALS patient becomes the first power user of a speech implant - source

(9:13) - Sign-off

Washington forces Anthropic to disable Fable 5 and Mythos 5, new data quantifies the cost of cleaning up AI-written code, and researchers reveal a backdoor that hides between chip and model.

(0:00) - Intro

(0:24) - U S forces Anthropic to pull Fable 5 and Mythos 5 - source

(1:10) - Why AI hasn't replaced software engineers - source

(1:54) - Senior engineers are cleaning up AI-generated code - source

(2:40) - The AI layoff wave is becoming a powder keg - source

(3:30) - HAMLOCK: a backdoor that hides between chip and model - source

(4:16) - curl pauses vulnerability reports for July - source

(5:02) - Berkadia breach exposes 305,000 accounts - source

(5:43) - Microsoft's Wi-Fi office check-in draws privacy pushback - source

(6:28) - Command-injection flaw in the GD Perl module - source

(7:17) - Rio's 'homegrown' LLM looks like a model merge - source

(8:02) - Sign-off

Washington pulls Anthropic's newest models from foreign access, OpenAI draws a multistate probe, CISA sets a three-day patch clock, and npm finally kills automatic install scripts.

(0:00) - Intro

(0:23) - U S orders Anthropic to cut foreign access to its top models - source

(1:12) - Reporting points to Amazon as the trigger - source

(2:00) - OpenAI draws a multistate attorneys general probe - source

(2:48) - K P M G pulls an A I report built on hallucinations - source

(3:35) - Meta unwinds its two billion dollar Manus deal - source

(4:24) - C I S A sets a three-day deadline for a critical Ivanti flaw - source

(5:10) - N P M 12 turns off automatic install scripts - source

(5:59) - Why you should not trust large context windows - source

(6:45) - Python in the browser gets first-class PyPI packages - source

(7:28) - Sign-off

Inside the FBI's fake town built to be hacked, ShinyHunters' university-wide Oracle zero-day spree, Google's lawsuit over Gemini-built scams, and the world's first trillionaire.

(0:00) - Intro

(0:22) - The F B I built a fake town to get it hacked - source

(1:12) - ShinyHunters tears through universities via an Oracle zero-day - source

(1:57) - Attackers exploit a Check Point V P N authentication bypass - source

(2:42) - Labcorp settles the American Medical Collection Agency breach for thirty-five million - source

(3:33) - Google sues a network that used Gemini to build scams - source

(4:20) - A popular Perl password library shipped weak defaults - source

(5:08) - The F C C wants to end anonymous phone activation - source

(5:54) - SpaceX goes public and Musk becomes the first trillionaire - source

(6:42) - Talking to an A I model about your documents, by voice - source

(7:30) - A satirical take on A I's circular accounting - source

(8:19) - Sign-off

Claude Fable's relentless initiative, Apple yanking Siri's AI from Europe, and a record Bezos raise, set against a fast-moving security beat from Ivanti and Oracle to a 336-million-euro laundering takedown.

(0:00) - Intro

(0:24) - Claude Fable, Relentlessly Proactive - source

(1:11) - Apple Pulls Siri's AI From the EU - source

(1:59) - Bezos's Prometheus Raises $12B - source

(2:47) - Coinbase Lets AI Agents Spend - source

(3:38) - Deezer Hunts AI Music - source

(4:24) - Ivanti Sentry Exploited in a Day - source

(5:09) - Oracle Zero-Day Hits 100+ Orgs - source

(5:54) - 400 Arch Linux Packages Backdoored - source

(6:35) - Crypto Laundering Ring Falls - source

(7:20) - 1.5 Million Malicious Domains - source

(8:09) - Sign-off

Anthropic walks back a hidden Claude policy, Google ships a fast diffusion language model, and the security world runs hot - from a 67-million-record leak in Thailand to a rare arm64 kernel escape.

(0:00) - Intro

(0:28) - Anthropic walks back a hidden Claude safeguard - source

(1:21) - Google ships DiffusionGemma - source

(2:13) - Claude versus Codex: steer or dispatch - source

(3:00) - Datadog veterans launch Niteshift - source

(3:48) - Apple details Google's role in Siri - source

(4:37) - OWASP: prompt injection still dominates agentic AI risk - source

(5:22) - Rare guest-to-host escape in the Linux kernel on arm - source

(6:10) - Thailand probes a 67-million-record leak - source

(6:56) - FBI seizes sites in a clearance-targeting scheme - source

(7:49) - Most people can no longer spot AI content - source

(8:32) - Sign-off

Anthropic puts its first Mythos-class model in public hands as Claude Fable 5, Apple defends an AI privacy bargain struck on Google's servers, and Microsoft ships a record two-hundred-six-patch Tuesday.

(0:00) - Intro

(0:21) - Anthropic ships Claude Fable 5, its first public Mythos-class model - source

(1:11) - Simon Willison's hands-on verdict on Fable 5: a beast - source

(2:01) - Ethan Mollick: working with Mythos feels like being a patron, not a wizard - source

(2:56) - Apple says its AI stays private even when it runs on Google's servers - source

(3:43) - EU forces Meta to host rival AI assistants on WhatsApp for free - source

(4:27) - Patch Tuesday hits a record 206 CVEs, and AI gets the blame - source

(5:16) - A single stray character in the Linux kernel opened a path to root - source

(6:04) - Russian groups are still exploiting a year-old WinRAR flaw against Ukraine - source

(6:54) - Waymo open-sources a model of how humans avoid crashes - source

(7:43) - Three AI labs, three theories of superintelligence, one looming memo - source

(8:29) - Sign-off

OpenAI joins Anthropic in filing for an IPO, Apple rebuilds Siri on a Gemini-powered architecture, and a wave of supply-chain attacks targets developers' AI coding tools.

(0:00) - Intro

(0:26) - OpenAI files for IPO, following Anthropic - source

(1:16) - Apple rebuilds Siri on a Gemini-powered architecture - source

(2:05) - Microsoft open-source packages laced with credential stealer - source

(2:51) - Anthropic's Mythos can build exploits from patches in hours - source

(3:48) - Google patches an exploited Chrome zero-day - source

(4:32) - Meta asks court to hold NSO in contempt over WhatsApp attacks - source

(5:20) - Study flags security risks in Chinese AI coding models - source

(6:14) - Hades campaign poisons 19 PyPI packages - source

(7:06) - Pentagon lists Alibaba, Baidu, BYD, Unitree as backing China's military - source

(8:01) - Sign-off

The Miasma worm tears through Microsoft's GitHub repositories, a new wave of AI coding tools lands on the desktop, and Uber, Wayve, and Waymo line up for a robotaxi showdown in London.

(0:00) - Intro

(0:24) - A self-replicating worm turns cloning a repo into a compromise - source

(1:19) - Google's Colab CLI hands AI agents a cloud supercomputer - source

(2:13) - GitHub's standalone Copilot app becomes a control room for agents - source

(3:12) - Simon Willison's plugin lets agents edit text without mangling it - source

(4:08) - Dries Buytaert: agents abandon tools that cost too much to adopt - source

(5:05) - Uber burned its AI budget early, and the bill was a warning - source

(6:04) - OpenAI's Lockdown Mode for ChatGPT cuts off the exits - source

(7:01) - CISA: actively exploited SolarWinds Serv-U flaw, patch by June 19 - source

(7:55) - TechCrunch's midyear roundup of the worst breaches of 2026 - source

(8:53) - Uber, Wayve, and Waymo line up for a robotaxi showdown in London - source

(9:48) - Sign-off