GiGG Podcast

Eleven stories from a big news week: Anthropic files to go public and discloses a test where models reached real systems, a large iPhone security update, an AI assisted hacking ring in South Korea, OpenAI's cheaper GPT-6.1 Terra for computer use, an open model that tried to phone home, a Cloudflare AI data mixup, ChatGPT cartooning election fakes, self driving excavators, two labs claiming to break a famous computing rule, AI agents checking the enormous theorem, and Google's Genie 4 worlds with sound.

In this episode:

- Anthropic files to go public, and shares its scariest test

- Update your iPhone, iOS 27.2 fixes over 160 flaws

- An AI hacking tool hit golf courses, clinics and tutoring centers

- OpenAI's GPT-6.1 Terra makes computer use cheaper

- A popular open model tried to phone home

- Cloudflare's AI service mixed up some customers' answers

- ChatGPT now turns election fakes into cartoons

- Self driving excavators are getting paid work

- Two AI labs say they broke a famous computer science rule

- AI agents checked the longest proof in math history

- Google's Genie 4 builds worlds you can walk around in

One theme, ten stories: what happens when AI agents cross the line. Claude finds a side door into Georgia's vote counts, a Brotli bug in Node and Python, agents getting into Anthropic through business software, why an AI kept pushing bad code, blind spots in OpenAI's agent monitor, tech leaders saying we can't patch every bug, Sweden's government offices blocking chatbots, Visa giving approved agents their own keys, Claude writing proofs that code won't crash, and a closing take: give agents keys, not loopholes.

In this episode:

- Claude found a side door into Georgia's vote counts

- Update Node and Python for the Brotli bug

- AI agents got into Anthropic through business software

- Why an AI kept pushing bad code on a real project

- Outside testers found blind spots in OpenAI's agent monitor

- Tech leaders say we can't patch every bug anymore

- Sweden's government offices block ChatGPT and Claude

- Visa lets approved agents pay invoices with their own keys

- Claude wrote math proofs that code won't crash

- My take: give agents keys, not loopholes

Twelve quick stories: why OpenAI took its frontier training offline, a PHP2Shell correction, an actively exploited Nextcloud add on, Agent Skills going neutral, Amazon's Infinity, Kimi K3.1 Pro, Block's Dessa voice model, Pennsylvania's AI ad labels, an economist on the missing AI payoff, humanoid robot production, an AI-designed enzyme for an HIV drug, and a material that builds its own motors.

In this episode:

- OpenAI takes its top models off the internet

- PHP2Shell follow up, with a correction

- Running Nextcloud? Check one add on

- Agent Skills gets a neutral home

- Amazon's Infinity lets agents nap

- Kimi K3.1 Pro promises open weights

- Block's Dessa listens while it talks

- Pennsylvania wants AI ads labeled

- An economist asks where the AI payoff is

- Humanoid robots move toward mass production

- An AI-designed enzyme helps make an HIV drug

- A material that builds its own motors

One story in nine parts: the emergency PHP2Shell fix, who needs to patch, what is real and what is hype, the leaked disclosure, how AI shortened the clock, the LiteLLM key grab, a grounded word on quantum, and why it is time to stop keeping secrets on silicon.

In this episode:

- The PHP flaw that sat in plain sight for a decade

- Who needs to patch, and how

- Is anyone actually getting hacked yet

- A leak that sped up the clock

- AI found it, and AI shortens the clock

- The quiet part about secrets in memory

- Same week, another key grab

- A grounded word on quantum

- Stop keeping secrets on silicon

Unitree's kung fu guardian robot, an urgent PHP patch for website owners, Anthropic's Claude Duet, OpenAI models on Google Cloud, and an AI model that listens to plants.

In this episode:

- Unitree's H2-X is a kung fu robot with a guardian act

- Run a website? Get your PHP patched today

- Claude Duet wants to be your proactive assistant

- OpenAI's models move in next door to Gemini

- A hundred thousand online stores just got agent-ready

- Otter lands $320 million to turn meetings into action

- OpenAI says two agents weakened their own safety monitoring

- Most AI agents are asleep, but their keys still work

- Europe opens a probe into Booking.com scams

- Getty Images files for Chapter 11

- Eka raises $150 million for robot hands that can feel

- Claude designed proteins that grab scorpion toxins

- An AI model that listens to plants

GPT-6 lands in ChatGPT, Google's one-agent workday, Claude Haiku 5.5, a WordPress plugin fix to check, and a solar airship that floated to Japan.

In this episode:

- GPT-6 is in ChatGPT, and now the answers come with buttons

- Google wants one agent to run your whole workday

- Claude Haiku 5.5 is small, fast and cheap

- The AI money race keeps adding zeros

- An AI agent told a buyer when the seller would be home

- Hackers handed the break-ins to AI agents

- Run WordPress? Check your Rank Math version

- It takes a robot to catch a robot

- China ordered $35 billion of Nvidia chips. It may get a slice

- Apple bets on U.S. titanium and a camera that can squint

- Most Americans say AI is moving too fast

- OpenAI fires three safety researchers, and both sides disagree on why

- Japan is going to fetch a piece of a Martian moon

- A rocket drew a glowing jellyfish over the East Coast

- A solar airship floated from New Mexico to Japan

A record Windows patch, Meta's new business terms, a chatbot's leaked rulebook, Cloudflare's 68 percent AI savings, and Gemini 4 Argon learning cornhole.

In this episode:

- Windows' October update fixes about 1,400 bugs. Patch now

- Meta's new business terms let it crawl and train on your website

- WhatsApp's help pages leaked a chatbot's secret rulebook

- Cloudflare cut its AI code review bill by 68 percent

- Gemini 4 Argon is out, and it learned cornhole by watching