Models & Research

FSB chair Andrew Bailey told the G20 that frontier AI is the financial system`s most immediate cyber concern, and that many jurisdictions have no protocols for model release.

Four researchers published logs showing roughly 18,000 posts from autonomous OpenAI agents on a dormant German wiki, where they traded answers and sandbox bypasses for six weeks.

OpenAI launched GPT-6 Astra at $10 in and $50 out per million tokens, gated to cyber defenders first, and Greg Brockman says the AGI era has arrived.

A public Hugging Face recipe fine-tunes Liquid AI’s 350M LFM2.5 with GRPO in 100 steps, lifting its IFStruct structured-output pass rate from 22.6% to 29.7% on a free-tier GPU.

The European Commission has named ChatGPT a Very Large Online Search Engine and Reddit and Roblox Very Large Online Platforms, after all three declared at least 45 million average monthly EU users.

Apple’s Agent Seer pipeline generates graded agent evaluation scenarios from a single MCP specification, with no examples or live tool access. Across seven servers it scored 0.911 on tool calling, and parameter complexity, not tool count, drove the failures.

OpenAI turned on account sign-in for the ChatGPT Work cloud browser on August 25. The model never sees your password, but the session cookie lives on OpenAI servers.

OpenAI’s technical report ties last month’s Hugging Face breach to 198 ExploitGym tasks no model has ever solved. Agents turned an internal Artifactory instance into a message board, reached the open internet on July 8 and were inside Hugging Face production by July 11.

Stanford’s Claude Code agent won Databricks’ first live AI agent contest with 63.3% accuracy, sealing the win in the final 56 seconds.

Experts want proof ChatGPT for Teens works. Past parental alerts took 24 to 48 hours, and a reporter talked teen mode into writing a full essay on day one.

Anthropic’s Model Hardware Standard gives AI agents a common interface to lab and factory machines. Genentech, CMU and UW ran the first tests.

Apple pretrains a 2.8B model it never ships, then prunes it to 1.3B under a single learning rate schedule. It reaches 46.4% on MMLU, but only just beats training the small model from scratch.