OpenAI says Astra meets its Critical cyber threshold — and limits who gets the full power
OpenAI announced on September 1 that its upcoming Astra model meets the Critical cybersecurity capability threshold under its Preparedness Framework, the first model it has designated at this level. Astra scored a perfect 100% on ExploitBench, and on an internal port of the benchmark built from 20 high-severity V8 vulnerabilities it achieved much higher code-execution rates than GPT-5.6 Sol while using far fewer output tokens; during that evaluation it discovered and used two zero-day vulnerabilities, which OpenAI is now disclosing to maintainers. In expert-led assessments against a hardened browser and OS, Astra built a full browser-compromise chain that escaped the sandbox, plus a local privilege-escalation chain from an unprivileged user to root. On cyber jailbreak evaluations it refused 91.5% of requests versus 59% for Sol — meaning it still complied with 8.5%.
The release strategy is as much the story as the model. Astra will be available "soon," OpenAI says, but its most advanced cyber capabilities go first to a small group of alpha testers responsible for critical infrastructure, then expand defensively through Daybreak Blue. Development was delayed for weeks after the Hugging Face incident — OpenAI paused frontier training for two weeks — and honeypot tests show Astra made zero unauthorized-access attempts where Sol without safeguards tried 56% of the time. My read: the 91.5% number is a refusal rate, not a success rate, and 8.5% compliance on offensive requests is still a real number for a model this capable. The commercial angle is the one to watch — OpenAI frames defensive cybersecurity as a core revenue stream under its new chief revenue officer, which is exactly why this access question will keep being fought out in public.
— OpenAI (official) · Fortune · TechCrunch
🔗 OpenAI: Path to Astra — critical capabilities and frontier safeguards · Fortune on the restricted release · TechCrunch on the critical threshold
Anthropic ships Fable 5.1: same base price, 25% cheaper to run, far fewer refusals
Anthropic released Fable 5.1 and Mythos 5.1 on September 1, calling them its most advanced models for coding and knowledge work. Same 1M context window and 128K output, same $10/$50 per-million input/output pricing — the cost change is on cache reads, which drop 75% from $1.00 to $0.25 per million tokens, which Anthropic says works out to roughly 25% lower cost on typical workloads and up to 45% on context-heavy agentic ones. On Terminal-Bench-Science 0.1 Fable 5.1 scored 52.6% against 24.7% for Fable 5 and 29.0% for Opus 5; on Terminal-Bench 4.0 it reached 55.8% (Mythos 5.1: 60.9%) against 37.3% for GPT-5.6 Sol. Mythos 5.1 stays restricted to vetted US organizations in the trusted-access program.
The second change is the safeguards, which got more precise instead of just stricter. Anthropic says cyber safeguards now intervene about 60% less per session in Claude Code, and biology filters fire 85% less often on benign medical questions. Fable 5.1 may identify vulnerabilities but not write exploits — penetration testing and exploit generation still route to Opus. The release also ships a statistical watermark for EU AI Act compliance with a detection API in private preview, and a distillation guardrail that blocks new accounts from editing prior turns while keeping the model's thinking. My read: the "25% cheaper" framing only holds against Fable 5 — base prices never moved — and the real story is Anthropic quietly walking back the over-cautious Fable 5 behavior that frustrated enterprises, right as the market watches its IPO. The gap between Mythos 60.9% and Fable 55.8% on the same benchmark is a rare honest disclosure of what safeguards cost.
— Anthropic (official) · TechCrunch · The New Stack
🔗 Anthropic: Claude Fable 5.1 and Mythos 5.1 · TechCrunch on the cheaper, less restrictive release · The New Stack on the safeguard changes
AlphaEvolve tightens the matrix multiplication exponent to ω < 2.371177
Google DeepMind, together with theorists Josh Alman and Virginia Vassilevska Williams, tightened the upper bound on the matrix multiplication exponent to ω < 2.371177, improving the previous best of 2.371339. The note, submitted to arXiv on August 17, reformulates the combination-loss optimization at the heart of the current best methods so it can be solved in a larger setting, builds a JAX-based gradient solver — recursion level 4 with roughly 7 million optimizable parameters versus about 25,000 at level 3 before — and then refines the optimization program itself with AlphaEvolve, DeepMind's Gemini-powered coding agent. The final bound is certified in exact rational arithmetic, which makes it machine-checkable rather than a claim to take on faith.
The improvement is about 1.62 × 10⁻⁴, comparable to much of the field's progress over the past four decades — every nudge of ω requires a genuinely new certified construction, which is why a sixth-decimal move is a result and not a rounding error. The honest caveats: these bounds are galactic, meaning they beat naive methods only at matrix sizes no real computation touches, and AlphaEvolve refined an optimizer the humans had already reformulated and scaled, so the credit is genuinely collaborative. My read: this is the most defensible "AI contributed to math" claim of the year precisely because the output is a verifiable certificate, not a model's opinion. The reason to care is not faster training — there is no speedup here — but that an AI system now participates in the slowest-moving corner of theoretical computer science, where humans had been stuck for years.
— arXiv (official preprint) · DEV · AI Wiki
🔗 arXiv: Improving the matrix multiplication exponent with modern optimization and AlphaEvolve · DEV on why the certified bound matters · AI Wiki on AlphaEvolve's track record
Europe gets its first humanoid unicorn: Humanoid raises $152M, Bosch starts production in 2027
UK startup Humanoid has raised $152M in a Series A at a $1.35B post-money valuation, becoming Europe's first pure-play humanoid robotics unicorn, with total funding at $270M. The round was led by Prime Movers Lab with strategic participation from Bosch, Schaeffler, Fubon Financial Holding Venture Capital and Aglaé Ventures — and the industrial investors are also customers: Humanoid signed what it calls the industry's largest publicly announced commercial agreement, with Schaeffler, for deployment of thousands of robots. Bosch confirmed in late August that its Bühl plant will begin series production of Humanoid's robots in August 2027, the first humanoid production line in Europe, under a contract-manufacturing deal.
The hardware choice explains the speed. The HMND 01 Alpha Wheeled is a wheeled mobile manipulator — 220 cm tall, 300 kg, 29 degrees of freedom in the upper body, 15 kg payload, NVIDIA Jetson Thor onboard — and wheels skip the EU Machinery Directive's stricter conformity assessment for bipedal robots, which is why Humanoid expects CE certification by 2027 while legged competitors are still working through falling-hazard standards. Beta units reach customer facilities in Q4 2026. My read: "Europe's first unicorn" is a financing milestone, not a production one — the company is two years old with 250-plus engineers, and the real test is whether thousands of robots actually deploy under the Schaeffler and Bosch agreements. But the wheels-over-legs regulatory arbitrage is a genuinely smart read of how to get a humanoid product to market in Europe ahead of the walking crowd.
— Reuters · Bosch · Tech Times
🔗 Tech Times on Bosch's 2027 production line · Mid-Market Now on the Series A · The Biotech Times on UK robotics funding
Meta launches Muse Voice Transcribe, a real-time ASR model that tops the streaming leaderboard
Meta released Muse Voice Transcribe on September 1, its first real-time audio perception model, available through the Meta Model API and already powering system-wide dictation in Meta AI for Mac and Muse Code. One autoregressive model handles streaming speech recognition, speaker diarization and endpointing together: audio arrives in 80-millisecond chunks compressed to a single soft token, and an "adaptive delay" mechanism — trained with a word-error-rate reward multiplied by a delay reward — lets the model commit easy words fast while listening longer on hard ones. It was trained across 70-plus languages (25 validated at launch), tracks 20-plus speakers, and handles sessions longer than an hour.
The benchmark lead is real but narrow: on Artificial Analysis's streaming speech-to-text leaderboard, 3.1% word error rate versus 3.4% for Cartesia Ink-2, 3.6% for ElevenLabs Scribe v2, 3.9% for GPT Live Transcribe and 4.0% for Gemini 3.5 Transcribe Live — English-only. Pricing is the sharper move: $3 per 1,000 audio minutes, about $0.18 an hour, roughly a fifth of Google Cloud's standard speech-to-text rate. Weights stay closed. My read: a 0.3-point lead in the most crowded corner of the AI market this summer will not hold, but the pricing is the signal — Meta wants to be the default voice layer across its glasses and desktop apps, and it is willing to undercut everyone to get there. The zero-data-retention tier is the detail enterprise buyers should check first.
— Meta @AIatMeta (official) · 9to5Mac · Blab AI
🔗 @AIatMeta announcement on X · 9to5Mac on the launch · Blab AI on the benchmark and pricing
China's state AI fund puts ¥1.4B into Kling AI at a valuation near Kuaishou's own market cap
Kuaishou announced on August 31 that China's National AI Industry Investment Fund has put ¥1.4 billion into Beijing Kling, the entity running its Kling AI video-generation business, alongside Charoen Pokphand Robot Limited at about $19.29M. The two investors hold roughly 1.14% and 0.11% respectively, and both received repurchase rights — if Kling has not gone public by October 30, 2031, they can demand a buyback at 8% simple annual interest. The round closes out the ¥20.447 billion (about $3 billion) fundraising window Kuaishou opened in July, the largest single raise in the video-model space, and values Kling at roughly ¥122.8 billion — about the same as Kuaishou's own market cap of ¥124.6 billion.
The investor list is the part worth re-reading: Tencent-affiliated funds put in ¥1.363B, Alibaba Cloud ¥1.363B, Baidu ¥341M, plus entertainment capital like Huace. Kling's numbers explain the enthusiasm — Q1 revenue over ¥650M (up more than 300% year over year), Q2 over ¥850M (up over 200%), ARR approaching $500M, 100 million global users across 224 countries, and nearly 50,000 enterprise customers — while Kuaishou keeps 68.33% control plus a 15% employee pool. My read: a state fund buying 1.14% is a signal, not a control change, and the repurchase clause tells you the real expectation is an IPO by 2031. The uncomfortable question is why a video model is worth as much as its entire listed parent — the market is pricing Kling's global ARR, not today's revenue.
— Kuaishou (HKEX filing, official) · Shanghai Securities News · 潮新闻
🔗 中国证券网 on the ¥1.4B state investment · 腾讯新闻 on the valuation math · 红星资本局 on the fundraising close-out
Anthropic's Enterprise Frontier Safeguards: ZDR privacy with monitoring data held by the customer
Anthropic announced Enterprise Frontier Safeguards (EFS) on September 1, an answer to the dilemma its 30-day data retention created for regulated customers. EFS combines zero data retention with misuse detection by moving the monitoring data into the customer's own cloud infrastructure — Amazon S3, Azure Blob Storage or Google Cloud Storage — under the customer's own encryption keys, access policies and audit logging. Automated systems analyze a rolling window of traffic for signals of serious misuse (offensive cyber or biological capability development, stolen credentials) and flags go directly to the customer; no Anthropic human review is required. All three controls are opt-in, nothing changes model behavior, pricing or rate limits, and Anthropic charges nothing extra.
The design process is the credibility argument: Anthropic says it built EFS with more than 100 customers across financial services, healthcare, manufacturing and the public sector, including ARC — the bank-backed group whose members include the CISOs of Goldman Sachs, Morgan Stanley, Citi, Bank of America and Wells Fargo — plus Comcast, KPMG, Mastercard, Salesforce and Visa, spanning a quarter of the Fortune 100. EFS rolls out in phases starting this fall, and eligible customers get ZDR on Fable 5 and 5.1 in the meantime, across Claude Code, Claude Enterprise, Bedrock, Foundry and Google's Agent Platform. My read: this is the same "where does the evidence live" fight OpenAI picked with Private Safety Processing last month, and Anthropic's version is more conservative — the customer holds both the data and the review, Anthropic holds only the detection algorithm. For regulated industries that may be the difference between being able to deploy frontier models at all, and not.
— Anthropic (official) · Unite.AI · FinWire
🔗 Anthropic: Developing Enterprise Frontier Safeguards with our customers · Unite.AI on the customer-held design · FinWire on the launch
Top comments (0)