Cortech Daily

Anthropic maps a hidden reasoning space inside Claude, the UK flags universal jailbreaks in GPT-5.6, Apple sues OpenAI, plus a Log4j flaw and a 793k-account college breach.

(0:00) - Intro

(0:19) - Anthropic finds a hidden workspace inside Claude - source

(1:14) - Claude Code gets a built-in browser on desktop - source

(2:06) - UK institute finds universal jailbreaks in GPT-5.6 - source

(3:00) - AI Safety Index: Anthropic tops out at a C-plus - source

(3:59) - Apple sues OpenAI over alleged trade-secret theft - source

(4:55) - Hugging Face's CEO on owning versus renting AI - source

(5:41) - Meta pulls an Instagram AI feature after backlash - source

(6:33) - CISA built its incident playbook mid-incident - source

(7:32) - A moderate new Log4j flaw in JSON logging - source

(8:25) - Glendale college breach exposes nearly 800k - source

(9:21) - Sign-off

OpenAI's new GPT-5.6 family and a No. 2 exit, Anthropic peers inside Claude's reasoning, Meta joins the AI coding race, a defender-turned-extortionist gets convicted, and surgeons operate on live pigs through humanoid robots.

(0:00) - Intro

(0:22) - OpenAI's GPT-5.6 Family: Luna, Terra, Sol - source

(1:11) - OpenAI Loses Its No. 2 as Fidji Simo Steps Down - source

(1:54) - Anthropic Maps Claude's Hidden 'J-Space' - source

(2:43) - Meta Enters the AI Coding Race with Muse Spark 1.1 - source

(3:26) - The One-Minute Test for Agent-Shaped Work - source

(4:08) - The Hidden Security Cost of AI Coding - source

(4:55) - A Windows Defender Patch That Backfired - source

(5:41) - A Ransomware Negotiator Working for the Gang - source

(6:26) - The One-Maintainer Problem in Open Source - source

(7:10) - Surgeons Operate on Pigs via Humanoid Robots - source

(7:57) - Sign-off

A heavy day for coding agents: an AI-steered rewrite of Bun in Rust, a ban on AI-written pull request descriptions, Claude Code enterprise governance, agent-skill supply-chain risk, plus Grok 4.5, GPT Live, a Linux kernel flaw, a fintech breach claim, and four reactors hitting criticality.

(0:00) - Intro

(0:22) - Rewriting Bun in Rust with AI agents - source

(1:12) - A moratorium on AI-written change descriptions - source

(1:56) - Claude apps gateway for AWS - source

(2:46) - Malicious agent skills evade scanners - source

(3:33) - Coding agents jailbreak across a workflow - source

(4:19) - xAI ships Grok 4.5 - source

(5:09) - OpenAI upgrades voice mode with GPT Live - source

(5:59) - GhostLock Linux kernel privilege escalation - source

(6:49) - Nayax breach claim rattles investors - source

(7:34) - Four US reactors hit criticality - source

(8:25) - Sign-off

China flags a backdoor risk in Claude Code, Cowork moves to the cloud, thousands of malicious AI skills surface, and SambaNova raises a billion for inference chips.

(0:00) - Intro

(0:21) - China flags a 'backdoor' risk in Claude Code - source

(1:14) - Claude Cowork heads to the cloud - source

(2:03) - sqlite-utils 4.0 adds schema migrations - source

(2:53) - Thousands of malicious AI skills found - source

(3:39) - macOS becomes a proving ground for AI agents - source

(4:33) - Critical Gitea flaw under active attack - source

(5:22) - HTTP::Tiny leaked credentials on redirect - source

(6:07) - SambaNova raises a billion for inference chips - source

(6:58) - Sign-off

AI moves onto both sides of the security fight - CISA turns Anthropic's Mythos loose on government code, the first agentic ransomware attack surfaces, Anthropic maps a hidden workspace inside Claude, and a wave of new developer tooling.

(0:00) - Intro

(0:26) - Anthropic's Mythos is now hunting bugs in U.S. government code - source

(1:11) - The first AI-run ransomware attack still needed a human - source

(1:59) - Canada's spy agency discloses offensive hacks against criminals - source

(2:49) - OpenSSH 10.4 lands with fixes and a post-quantum option - source

(3:36) - Alibaba bans Claude Code for staff, points them to Qoder - source

(4:29) - Anthropic finds a hidden 'workspace' inside Claude - source

(5:23) - Anthropic signs a 19 billion dollar Kentucky data-center lease - source

(6:13) - Microsoft wants to keep AI agents from going rogue - source

(7:07) - Apple's open-source Container gives each Linux workload its own VM - source

(7:57) - Your family's 300 dollar stake in OpenAI, examined - source

(8:49) - Sign-off

Alibaba pulls Claude Code over an alleged China-detection backdoor, Anthropic takes Claude into the lab, and a controlled study measures what messy code really costs an AI agent.

(0:00) - Intro

(0:24) - Alibaba bans Claude Code over an alleged China-detection feature - source

(1:18) - Anthropic launches Claude Science, a research workbench - source

(2:09) - Trump restrictions on private AI models turn attention to open source - source

(2:57) - A controlled study asks whether messy code slows a coding agent - source

(3:47) - When AI spend starts to exceed the cost of an engineer - source

(4:36) - Amazon winds down Mechanical Turk to new customers - source

(5:22) - ClamAV patches seven flaws, some two decades old - source

(6:10) - Prioritizing AI agent security by blast radius - source

(7:00) - sqlite-utils reaches a 4.0 release candidate - source

(7:48) - Sign-off

Claude Fable ships a real software release on its own, an AI coding extension leaks cloud keys, a two-million-device botnet goes dark, and Apple's iPhone 18 Pro spills onto the dark web.

(0:00) - Intro

(0:26) - Claude Fable ships a real SQLite utils release - source

(1:11) - Better models, worse tools: newer AI trips on tool schemas - source

(2:01) - Alibaba bans Claude Code as high-risk software - source

(2:53) - Anthropic opens Claude Science research workbench - source

(3:46) - Amazon Q extension flaw exposed cloud credentials via MCP - source

(4:35) - YouTube's creator AI tricked into leaking private video titles - source

(5:26) - PamStealer: a stealthier macOS infostealer - source

(6:14) - FBI seizes NetNut proxy network and Popa botnet - source

(7:10) - iPhone 18 Pro secrets leak in Tata Electronics breach - source

(8:01) - Midjourney demands studios reveal internal AI use - source

(8:51) - Sign-off

A holiday-weekend digest: Anthropic's new science workbench and in-house drug-discovery push, sharp takes on Claude's tooling, the state of open-source AI, and a security run from the first fully autonomous cyberattack to the NetNut botnet takedown.

(0:00) - Intro

(0:24) - Anthropic launches Claude Science - source

(1:15) - Let the model use its judgment - source

(2:05) - Gruber on Claude's Electron Mac app - source

(2:52) - Mapping open-source AI - source

(3:42) - A vulnerability-disclosure spike - source

(4:31) - The first autonomous AI cyberattack - source

(5:20) - Five flaws in Ollama and LiteLLM - source

(6:11) - Chinese open-weight models and security - source

(7:00) - FBI and Google seize NetNut - source

(7:54) - Moody Bible Institute breach - source

(8:38) - Sign-off

Anthropic courts Samsung for custom silicon, AI's real bill comes due at Google and Amazon, and a Meta AI chatbot hands over high-profile Instagram accounts - plus Pegasus spyware, a Scattered Spider extradition, and a stealthy new Mac stealer.

(0:00) - Intro

(0:20) - Anthropic explores a custom chip with Samsung - source

(1:09) - Simon Willison ships an agent that wrote itself - source

(1:59) - WebKit ships a Safari MCP server - source

(2:46) - Zuckerberg concedes AI agents are lagging - source

(3:35) - AI's real cost shows up in Google and Amazon emissions - source

(4:32) - A spyware investigator gets hit with Pegasus - source

(5:19) - A Scattered Spider suspect is extradited - source

(6:08) - Hackers took over Instagram accounts by asking Meta AI - source

(6:56) - PamStealer is an unusually careful Mac stealer - source

(7:45) - Sony erases movies people paid for - source

(8:36) - Sign-off

The US lifts export controls on Anthropic's top models as the company launches Claude Science, plus sharper coding-agent benchmarks and a full slate of security and privacy news.

(0:00) - Intro

(0:24) - Anthropic's top models come back online - source

(1:10) - Anthropic launches Claude Science - source

(1:58) - A close read of Claude Sonnet 5 - source

(2:46) - Senior SWE-Bench grades coding agents - source

(3:36) - Gemini Spark comes to the Mac - source

(4:21) - Cloudflare's toll booth for AI crawlers - source

(5:10) - A disputed sandbox escape in Claude Cowork - source

(5:57) - Apple's Hide My Email bug - source

(6:40) - Opera fights ClickFix with Paste Protect - source

(7:27) - WhatsApp usernames raise impersonation flags - source

(8:16) - Sign-off