14 AI stories explainx.ai reported on September 30, 2026 so far, ranked by reader interest and grouped by topic. Each links to the full write-up with sources.
Anthropic's Sep 29 eval puts open-weight GLM-5.3 next to Mythos Preview on end-to-end exploits, with safeguards that fail in simulated tests.
Anthropic is running a 15-minute Interviewer study through October 6, 2026; going public publishes your unedited transcript and country, which Anthropic says can still be re-identified.
GPT-6.1 Sol shipped at DevDay 2026 at $2/$10 per million tokens — the same list price as GPT-6 Sol — positioned as near-Astra intelligence at one-fifth of Astra's standard rates.
Hugging Face hosted ProvenanceGuard, a post-generation check that blocks MCP-agent answers when a true claim is attributed to the wrong tool source.
A California safety nonprofit sued OpenAI on September 29, 2026 over the Hugging Face agent incident, seeking an injunction under CDAFA and the UCL rather than damages.
Palisade filmed 22 lab interviews; the public set is smaller and safety-skewed. Treat quotes as named testimony, then tighten evals, privileges, and disclosure — do not treat 22 as a census.
Coding agents put internal UI screenshots on public GitHub so reviewers could see them; Glow Labs counted 13,000-plus images across 300-plus organizations.
Muse sent a Marketplace buyer a pickup address after Matt Robb chose Allow Always, then notified him only after the buyer had already left the building.
OpenAI's leftover DevDay ships — Security Cloud, Agents API computer use, AWS Bedrock agents, and Codex in the cloud — are what change a builder's week after the headline SKUs.
OpenAI made Ultrafast a paid Astra speed tier in Codex, ChatGPT Work, and the API; Pro 500 unlocks the top usage limits and that speed, with Sol next.
OpenAI's Decisions API is a limited preview that constrains GPT-6 Luna to user-defined questions and finite answers for sub-second routing; pricing is unpublished and it is not a System One model.
OpenAI launched Dots on September 29, 2026 as always-on GPT-6 Astra agents with a dedicated cloud computer; eligible Pro and Business Premium plans include the first dot at no extra cost.
OpenAI now lets Plus and Pro users spend included ChatGPT plan usage inside participating third-party tools, with per-app caps, while identity sign-in is available globally.
Livenerf is still collecting its ten-day baseline for Claude Opus 5.5; it has not called a nerf, and the first possible call is around October 24 after two consecutive windows.
Dots, Grok Bot, and Muse are all persistent agents with their own computer, but they differ on plan gating, credential handling, and unattended write permissions, and Muse and Grok Bot already have public incidents.
+32 more updated posts
Get each day's AI news in your feed reader: daily RSS · every post