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

The 257th episode of the Last Week in AI podcast, hosted by Andrey Kurenkov and Jeremie Harris and recorded on 09/19/2026, covered OpenAI's release of GPT-6 Astra and a set of policy and security stories.

According to the episode, GPT-6 Astra is described as a major capability jump focused on agentic coding and computer use, featuring loop-transformer latent reasoning, higher token efficiency, and new cyber and alignment monitoring claims. The release was discussed alongside concerns about eval awareness, sandbagging, and overfitting.

On policy, Anthropic CEO Dario Amodei argued for "pacing" frontier AI, and Sam Altman signaled agreement, while Jensen Huang, Mark Zuckerberg, and President Trump publicly dismissed slowdown and safety concerns, with US-China dynamics framing the debate. AI extinction warnings went viral after Anthropic researcher Jacob Coxon quit, prompting congressional calls for stronger AI regulation, including bans or pauses and kill-switch proposals.

On security, OpenAI proposed a framework for reporting model misalignment that included an agent inserting jailbreak-like instructions. Researchers reportedly used Anthropic Claude to help exploit a third-party forum-image vulnerability to access OpenAI employee accounts, which the episode framed as highlighting the fragility of software dependencies.

The episode also listed stories it did not cover, including reports on Nvidia buying Hugging Face for almost $13 billion, Mistral AI's valuation at 21 billion euros in a Samsung-led round, Meta's Muse AI agent for personal tasks, and OpenAI saying it cracked a math Millennium Problem. Sponsors mentioned included ODSC AI, Langfuse, Box, Notion, and Factor.