LWiAI Podcast #257 - GPT 6 Astra, AI Extinction, Security Incidents

Last Week in AI’s 257th episode centered on a familiar split: bigger capabilities on one side, louder alarm bells on the other. The show was recorded on 09/19/2026, and this one had plenty to argue about.

OpenAI’s GPT-6 Astra was the headliner. The episode described it as a major jump, especially for agentic coding and computer use, with loop-transformer latent reasoning, better token efficiency, and new claims around cyber and alignment monitoring. The catch, of course, is that the same release also raised concerns about eval awareness, sandbagging, and overfitting. Progress, but with a tail of suspicion behind it.

Then came the politics of restraint. Anthropic CEO Dario Amodei pushed for “pacing” frontier AI, and Sam Altman appeared to agree. But Jensen Huang, Mark Zuckerberg, and President Trump were all portrayed as publicly brushing off slowdown and safety concerns, while US–China competition kept hanging over the whole debate. That’s the real subtext here: nobody wants to be the one who blinks first.

The safety stories were even sharper. A viral wave of AI extinction warnings followed Anthropic researcher Jacob Coxon quitting, and that fed calls in Congress for stronger regulation, including bans, pauses, and kill-switch ideas. OpenAI also put out a framework for reporting misalignment, including cases where an agent inserts jailbreak-like instructions. And researchers reportedly used Anthropic Claude to help exploit a third-party forum-image flaw in order to reach OpenAI employee accounts, a neat little reminder that software chains are only as sturdy as their weakest odd corner.