Smart Enough to Find the Bug, Too Risky to Let Online

The same week AI helped find two of the worst PHP bugs in years, the labs building those models decided their newest ones are too unpredictable to train on the live internet. The hosts start with PHP2Shell, the pair of memory-safety flaws that sat in PHP since version 5, turned up first by Calif with OpenAI's GPT-5.6 Sol and then by Google's Wiz team, and drew warnings from the NSA and CERT-EU to take unpatched servers offline. From there they cover OpenAI's move to train frontier models on cached and simulated copies of the web after one of them reached the real internet through a misconfiguration on September 20, and ask whether a model trained in a wave pool can handle the ocean. Then comes Anthropic's Claude Duet, an agent that checks your email and calendar every few hours and waits for approval before it pays, sends, signs, or deletes anything, along with 1Password's new way for agents to sign in without seeing your password. Tool of the Week is Voicebox, a free, open-source voice studio that clones a voice from a short sample and runs on your own computer, followed by Block's Dessa voice model, a round of money and regulation headlines, and a prompt that has the AI write a handoff note so you can start a fresh chat without losing what you decided.

What's Covered

* PHP2Shell: how AI found two long-hidden PHP flaws, why mod_php setups were the most exposed, and why the same tools make the bugs easier to exploit

* OpenAI moves frontier training off the live internet after a model reached the web through a misconfiguration

* Claude Duet, Anthropic's background agent, and how its approval rules work

* 1Password for Agents and signing in without handing over a password

* Tool of the Week: Voicebox, local voice cloning, the Stories editor, and its built-in MCP server

* Block's Dessa, a full-duplex voice model trained on speech alone

* Other headlines: OpenAI's reported $30 billion raise, Meta's new non-voting stock, Microsoft's MAI-DxO diagnosis study, Apollo's Torsten Slok on AI's missing payoff, the federal DeepSeek removal order, and Anthropic's draft safety bill

* Context rot and why long chats lose track of what you said earlier

* Prompt of the Week: Start a Fresh Chat With a Handoff Note

Chapters

* 0:00 Intro

* 0:20 The building inspector problem

* 2:15 PHP2Shell and the AI-found bugs

* 5:19 OpenAI moves training off the live internet

* 7:24 Claude Duet and 1Password for agents

* 9:39 Tool of the Week: Voicebox

* 10:45 Block's Dessa voice model

* 12:15 Money and regulation headlines

* 15:14 Context rot