OpenAI’s Astra Crosses ‘Critical’ Cyber Threshold After Finding Zero-Days Autonomously
OpenAI has disclosed that its upcoming Astra model is the first AI system to exceed the ‘Critical’ cybersecurity threshold under the company’s Preparedness Framework — a designation that signals the model can independently discover and exploit zero-day vulnerabilities across well-defended systems.
During benchmark testing on ExploitBench, Astra scored a perfect result. More striking still: while running a modified version of the test, the model autonomously discovered and exploited two previously unknown zero-day vulnerabilities in Chrome’s V8 JavaScript engine — without being instructed to look for them.
The ‘Critical’ classification means OpenAI must implement additional safeguards before any public release. The company said it paused Astra’s development in August to strengthen safety controls, and initial access will be restricted to vetted security researchers and approved testers. Astra has been trained to refuse harmful requests, with monitoring systems that flag and halt unauthorized behavior.
The announcement raises fundamental questions about AI-powered offensive cybersecurity capabilities. If a model can stumble upon zero-days during routine testing, what happens when similar systems are deliberately pointed at critical infrastructure?
Sources: TechCrunch, SecurityWeek
Anthropic Launches Fable 5.1 With Slashed Costs and Fewer Guardrails
Anthropic debuted Claude Fable 5.1 and its trusted-access twin Mythos 5.1 on September 1, positioning them as the company’s most capable models to date. The headline number: cache reads now cost just $0.25 per million tokens, a 75% reduction from Fable 5, which Anthropic estimates cuts typical workload costs by about 25% and highly agentic workloads by up to 45%.
Beyond pricing, Fable 5.1 brings meaningful performance gains. Self-reported Terminal-Bench-Science scores jumped from 24.7 (Fable 5) to 52.6, and the model outperforms both Opus 5 and OpenAI’s GPT-5.6 Sol across multiple benchmarks. Anthropic also reduced false-positive content restrictions, making the model less likely to refuse legitimate requests.
Mythos 5.1 is functionally the same model but without blocks on sensitive cybersecurity and biology research — access is limited to vetted organizations. Fable 5.1 is available now across Claude.ai, Claude Code, Cowork, and through the API on AWS, Google Cloud, and Azure.
The launch coincided with Anthropic sealing a $35 billion cloud-computing deal with Lambda, the Nvidia-backed cloud provider, for a Texas data center developed by Hut 8 covering roughly 350 megawatts of capacity. This deal is part of a broader infrastructure blitz that includes agreements worth $45B with Nscale and $50B with Fluidstack.
Sources: TechCrunch, SiliconANGLE, Bloomberg
Google Ships Gemini 3.8 Flash — Its Third Flash Model in Six Weeks
Google DeepMind released Gemini 3.8 Flash on September 2, its fourth Flash model in under four months and third in just six weeks. The relentless cadence is clearly aimed at keeping Google competitive in the fast-moving model race.
The numbers are solid: Gemini 3.8 Flash scores 59 on the Artificial Analysis Intelligence Index, up 3 points from 3.7 Flash, and approaches the performance of much more expensive frontier models on complex reasoning tasks. Pricing holds steady at $0.75 per million input tokens and $3.75 per million output tokens.
Google also launched Gemini 3.8 Flash Cyber, a specialized variant with vulnerability detection and automated patching capabilities, though access is restricted to participants in Google’s Fairwind security research program. Additionally, Google debuted Google Pics, a new image generation tool.
Sources: Artificial Analysis, The Register, 9to5Google
Sony and Warner Sue Anthropic Over ‘Blatant Theft’ of Song Copyrights
Sony Music Publishing and Warner Chappell Music filed a 48-page federal complaint against Anthropic in Northern California, naming the company and co-founders Dario Amodei and Benjamin Mann personally. The publishers call it "one of the largest and most blatant ongoing thefts of intellectual property in history."
At the heart of the complaint: Anthropic allegedly used approximately 20,000 copyrighted songs — including works like Eye of the Tiger, Hallelujah, Uptown Funk, Taylor Swift’s Paper Rings, and Mariah Carey’s All I Want for Christmas Is You — to train its Claude models without licensing or authorization. The publishers claim Anthropic sourced lyrics through piracy platforms including Library Genesis and the Pirate Library Mirror, and also scraped licensed lyric sites like Musixmatch and LyricFind.
The lawsuit seeks statutory damages of up to $150,000 per infringed work, plus $25,000 per instance of alleged copyright management information removal. The complaint leans heavily on Anthropic’s $1.5 billion settlement with book authors last September over similar training-data sourcing practices.
Anthropic has said it intends to defend itself vigorously in court.
Sources: TechCrunch, Axios, Fortune
Apple Enters the John Ternus Era — Without a Frontier AI Model
John Ternus officially became Apple’s chief executive on September 1, succeeding Tim Cook, who moves into the role of executive chairman. Ternus, a 25-year Apple veteran who led hardware engineering, now helms the only major technology company without a frontier AI model of its own.
His first big test comes on September 9, when Apple is expected to unveil its latest iPhone alongside a foldable device. The question on everyone’s mind: will Ternus accelerate Apple’s AI ambitions, potentially acquiring or developing a competitive LLM, or double down on the company’s strategy of integrating third-party models into its ecosystem?
Sources: Al Jazeera, TechCrunch