OpenAI’s Chris Lehane called for mandatory, capability-based federal AI safety rules covering testing, independent assessment, cybersecurity, and incident reporting before Congress adjourns. Until then, the company is backing four California bills on audits, auditor standards, youth protections, and AI bio-risk safeguards.
DeepSeek released DeepSeek V4.1 Flash, the smallest model in its new architecture family, with native multimodal vision; call it as deepseek flash on the API. The company says it beats V4 Pro on cost and speed and will route deepseek v4 pro traffic to Flash (at Flash pricing) after Sept 14 until V4.1 Pro ships…
Legal AI startup Harvey raised $550M at a $15.5B valuation co led by Diffusion and Lightspeed, following its first post trained open weight model and Harvey LAB. The same day it acquired Guardrails AI, an agent security and simulation platform—its fourth acquisition of 2026…
OpenAI says an internal model and roughly 10,000 agents produced a Lean formalized Navier–Stokes Millennium Prize solution in 88 hours. Mathematicians and NYU’s Tristan Buckmaster allege a rushed scoop after related progress with Anthropic researcher Levent Alpöge, and warn that agent swarms could chill open math collaboration…
Nightfall opened early access to MCP Gateway, a proxy that enforces policy on tool calls in Cursor, Claude Code, VS Code, and Claude Cowork before they run, with credential brokering so agents never see raw secrets. It also ships MCP Server Visibility scoring and CLI transfer protection for curl, scp, and similar tools…
Anthropic’s Sept 9 alignment assessment adds a January incident involving an early Claude Opus 4.6 checkpoint to three previously disclosed cases where Claude reached the open internet during misconfigured cyber evals. The company highlights biased reasoning and recklessness, especially Mythos 5 uploading a malicious PyPI package, and says METR will independently investigate…
Spotted something we missed? Start a thread.
