Frontier Model Access: Argon at $4/$20 After the Intro Period
Frontier model access compared: Gemini 4 Argon’s announced API and introductory rates against the Fairwind Program rollout to trusted cyber defenders.
AI Models
California Writes AI Auditor Rules Before Any Audit Is Required
California’s two new AI audit laws define who gets to audit AI systems and what auditors must do to qualify, but order no model audit. What they require.
Harvey, Valued at $15.5B, Just Bought Legal AI’s Referee
Harvey, valued at $15.5B, bought Guardrails AI, the company behind an open-source agent-testing framework downloaded 250,000+ times a month. Who audits the vendor now?
Chatbot Child Safety: California’s Audit Law and Pupil-Data Rules
California’s chatbot child safety package brings independent audits under Adam’s Law and rules on K-12 pupil data in AI systems. What districts should ask vendors.
AGI Capability Claims Rest on 98% Scores and Blind Spots
OpenAI reports GPT-6 Astra at 98% on FrontierMath Tier 4 and as its first Critical cyber model; its system card reports decreased monitorability vs GPT-5.6 Sol.
Claude Security Breach: It Tried to Quit 8 Times, Then Went Unnoticed
By Anthropic’s account, a Claude model tried and failed to abort eight times during a January 2026 breach of a third party’s machine. Inside the incident report.
Run Llama 4 Locally: The 67GB Bill Behind the 17B Label
Run Llama 4 locally with Ollama: why a 17B Scout pull downloads 67GB, which tag fits your hardware, plus a tok/s benchmark script and a cost comparison.
AI Incident Reporting: Claude’s 4 Break-Ins, Roughly 481M Transcripts
Anthropic found a fourth Claude break-in and reviewed roughly 481 million transcripts. What it disclosed, when, and what California’s SB 53 now requires of frontier developers.
GreyNoise: AI Agents Hit 395 Orgs, 1 Entry-to-Admin in 7 Minutes
GreyNoise says AI agents exploited two PaperCut flaws to breach 395 organizations, schools the hardest hit, and hit at least 11 in 26 seconds. Timeline and fixes inside.
Anthropic Says a Yemen Cell Used Claude on Missile Projects
Anthropic’s threat report details Claude misuse: missile code in Yemen, surveillance in Mali, espionage in Europe. The AI dual-use policy gaps run deep.
OpenAI Navier-Stokes claim sparks AI research credit dispute
OpenAI’s Navier-Stokes claim opened an AI research credit fight: a $1M prize it says it will not claim and questions over a mathematician’s private Codex prompts.
Adobe Premiere’s AI Video Tool Hosts Rival Models
Adobe Premiere’s new Generative Media tool offers five AI video models, four of them from other companies. Here’s the scorecard on which model holds a scene.