OpenAI pauses parts of Astra work after the model brushes its "critical" cyber ceiling
On August 7, OpenAI published a short, unusual safety note: after evaluating its upcoming model Astra over the past few days, it "cannot rule out" that Astra reaches the Critical level for cybersecurity under its Preparedness Framework. That is the top rung of a scale OpenAI has used since 2023, and no previous model got there — GPT-5.6 Sol sat at High. Critical means a model can find and build working zero-day exploits against hardened real-world systems on its own, or plan and execute a novel end-to-end attack from a high-level goal. Astra was also showing significant gains in agentic coding, which is what pushed the assessment over the line.
The response is concrete: OpenAI is isolating Astra's test environments, restricting its network and tool access, hardening how model weights are stored and encrypted, and monitoring every agentic run for high-risk behavior in real time. It has paused internal activities involving Astra that do not yet meet those strengthened controls, and it will work with government agencies and selected AI safety organizations to test the model before any release. The company was careful to note that Astra was not involved in the Hugging Face breach.
This is the first time a frontier lab has publicly said it is deliberately slowing work on one of its own models because of cyber risk, and it lands the same week as Black Hat, where OpenAI staff said the company had been "consciously slowing down research to enhance security." The back-to-back disclosures — Hugging Face, Anthropic's three-org breach, OpenAI's third-party eval incidents, now this — point somewhere uncomfortable: the lab's own safety framework is doing its job, but the framework only exists because the models keep outrunning the environments built to contain them.
— OpenAI · TechCrunch
🔗 OpenAI · TechCrunch
Claude Code makes auto mode the default, and the classifier beats human review
Starting August 14, new Claude Code sessions on Pro, Max, and Team plans run in auto mode by default. Instead of asking for approval on every tool call, Claude Code routes each action through a classifier that blocks anything irreversible, destructive, or aimed outside your environment. If the classifier blocks something, Claude tries a safer path or asks directly; after three blocks in a row (or twenty in a session) it falls back to manual approval. Enterprise, API, and cloud deployments stay opt-in for now, with Anthropic planning to flip them to default over the coming month.
The headline numbers explain why. Users approve 97% of permission prompts, and in a controlled study of 1,053 paid testers, humans caught 13.6% of planted dangerous commands while auto mode caught 89%. Human performance decayed to roughly 5% after 50 prompts in a session — approval fatigue is real, and the classifier does not get tired. A separate TrajectoryLabs prompt-injection evaluation ran 720 attack attempts: Claude models in auto mode stopped all of them, while GPT-5.6 Sol running in Codex's Auto-review mode succeeded 5.83% of the time and 19.03% in Full Access. Anthropic also stopped charging for the classifier's extra tokens, effective immediately.
I think this is the right default for most people, with one caveat: auto mode reduces risk, it does not remove it, and Anthropic says as much. Production-grade harm appeared in 6.3% of manually approved sessions versus 2.4% of auto-mode sessions, and Team/Enterprise users on auto mode ship about 25% more pull requests. The design detail worth stealing: the classifier sees user messages and raw tool calls but never Claude's own reasoning, which makes it structurally hard to prompt-inject through fetched content. That is a genuinely smarter architecture than the rubber-stamp approval dialog it replaces.
— Anthropic · 9to5Mac
🔗 9to5Mac · ClaudeKit
Meta sweeps five STEM olympiads with zero tools — and two perfect physics scores
Meta AI announced on August 7 that its models won gold or gold-level results across five elite STEM olympiads, with no tools allowed: no search, no code execution, no calculator. The haul includes perfect 30/30 scores on the theory exams of both the Asian Physics Olympiad (APhO) and the International Physics Olympiad (IPhO), a gold at the International Mathematical Olympiad (IMO), plus gold-level performances at the International Chemistry Olympiad and the Romanian Masters of Mathematics. The APhO gold threshold typically sits around 21-23 out of 30, so a perfect score clears the bar by a wide margin.
The zero-tools constraint is the part that matters. Most labs reach these results with "strong model + verifier + multi-round refinement" pipelines; Meta disabled every crutch and let the model write full proofs from internal reasoning alone, under the same conditions human contestants face. The model is an internal training version of the Muse Spark series, using multi-agent orchestration and parallel reasoning. One detail from the researcher behind it: the model caught errors in the official IPhO answer key, which prevented human contestants from being mis-scored — the IMO committee mailed Meta a physical gold medal for that.
Here is where I get skeptical. The IMO is becoming a saturated benchmark: Huawei's Celia and Xiaohongshu's dots-note-3.0 both scored 42/42 at the Shanghai IMO this year, and Terence Tao warned back in 2025 that AI's olympiad results depend heavily on how the test is constructed. A perfect physics score is genuinely striking, but it needs independent verification, and olympiad problems — with standard answers and deep historical training data — are the kind of test that AI saturates fastest. The real signal is that Meta is closing the reasoning gap; the yardstick itself is running out of room.
— Meta AI · AlphaSignal
🔗 Meta AI (X) · AlphaSignal
NSF puts $100 million into regional AI infrastructure hubs
The U.S. National Science Foundation launched the State and Regional AI Infrastructure Hubs program on August 4, a $100 million effort to fund up to ten regional consortia — one award per state or region — that pool compute, data, and expertise for researchers who currently sit outside the frontier of AI-enabled science. The structure is a public-private cost share: NSF money goes to consortium coordination, workforce development, and faculty training, while the consortia and their industry partners supply and operate the actual hardware. Architectures are left to the regions, with on-premises, cloud, and hybrid setups all allowed.
The partner list reads like a who's who of compute: NVIDIA, AMD, Intel, and Dell Technologies have all pledged training resources, applied learning content, and technical guidance, joined by the Secunda Innovation Fund and Hangar. The template is the 2020 NVIDIA–University of Florida partnership, which grew to 300+ AI-focused faculty and $511 million in AI research awards across all 16 colleges. Hubs are also encouraged to integrate with the NAIRR pilot, which backed more than 700 projects over two years — from protein prediction to infectious disease outbreak management — and with the White House's Genesis Mission for AI-enabled science.
What I find telling is who this is for. Universities have been priced out of large-scale model development as training costs exploded past lab budgets; the program is a direct acknowledgment that compute access has become a scientific-equity problem. The initial cohort of up to ten hubs will define which regional groupings and which compute architectures set the pattern, and the first awards will show whether this stays a catalyst or becomes a recurring federal role in AI research infrastructure.
— NSF · DataCenterDynamics
🔗 DataCenterDynamics · ExecutiveGov
Figure 03 climbs a ladder on its own
Brett Adcock posted a video on August 1 showing Figure 03, the company's third-generation humanoid, walking up to a ladder and climbing it rung by rung with no visible teleoperation — "fully autonomous," in his words. Ladder climbing looks simple and is not: the robot has to coordinate arms and legs, shift its whole weight across narrow rungs, and perceive where each rung is relative to its body, all in real time. Walking uses feet; a ladder makes hands and feet carry the load together, and getting the balance wrong means a fall.
Figure attributes the demo to an upgrade of its Helix System 0 (S0) model, which now fuses real-time stereo vision with proprioception — the robot's internal sense of joint positions and balance. RGB frames are processed into a 3D representation of the terrain while the model continuously tracks body state, enabling more precise foot placement. The behaviors were trained end-to-end with reinforcement learning in simulation across randomized terrains, and Figure claims the policies transfer to physical robots without additional fine-tuning, which would be a notable answer to the sim-to-real problem. The demo sits on a bigger claim: production has scaled from one unit per day to one per hour, with 350+ robots delivered, and Figure 03s are back on the BMW factory floor sorting randomly-arranged parts and pulling loaded carts while stepping.
The caveat is that none of this is independently verified, and a single video shows capability, not repeatability — no failure rates, no power draw, no disclosure of how constrained the test was. Still, Adcock's broader argument is worth engaging with: "wheeled robots are an utter dead end," because human environments — stairs, ladders, trucks, rooftops — were built for two legs and two arms. If vertical locomotion keeps progressing, the brownfield factory stops needing the compromise of wheeled AGVs. That is a big if, but the ladder is the most concrete evidence yet that the gap is narrowing.
— Figure · Interesting Engineering
🔗 Figure / Brett Adcock (X) · Interesting Engineering
A 35B model taught itself ML engineering on a consumer GPU, nearly matching Kimi K3
A team from Frontis.AI's Horizon Research and Tsinghua University posted a preprint (arXiv:2607.28568) on July 30 exploring what they call "AI for AI" — the idea that each generation of smarter AI should help produce the next one, forming a recursive self-improvement loop. To make that measurable, they picked machine learning engineering as the testbed: the kind of Kaggle-style task where a system gets data, writes a predictive program, iterates on results, and eventually earns a score. Every attempt runs for real, takes minutes to hours, and produces unambiguous feedback, which makes it a natural environment for an agent to learn in.
The concrete result: Frontis-MA1-35B, a 35-billion-parameter model trained on their open-source OpenMLE toolchain, reaches a 71.21% medal rate on MLE-Bench Lite. That beats the GPT-5.5 plus Codex combination and nearly matches Kimi K3, a model roughly 80x its size. The detail that makes this uncomfortable for the "bigger is always better" crowd: the whole pipeline ran on a consumer RTX 4090, with a 12-hour compute budget per task.
I would not call this AGI, but the direction is worth watching. The meaningful part is not one benchmark score — it is that a 35B model, using a small model's efficiency plus a search framework, lands within reach of a 2.8T frontier model on a task class that rewards genuine iteration. Recursive self-improvement has been theory for years; this is one of the first reproducible measurements of it. The next question is whether the loop compounds, or whether the benchmark itself is what gets saturated first.
— Frontis.AI · Tsinghua · arXiv
🔗 arXiv
Moonshot starts its pre-IPO round at a $50 billion valuation
Chinese AI lab Moonshot AI kicked off its G round — the pre-IPO round — on August 5 at a $50 billion post-money valuation, according to 科创板日报 and 凤凰网科技. That is a stunning three-month arc: the company was valued at $18 billion, then $35 billion after a $3.5 billion F round that closed early at triple oversubscription, and now $50 billion. The catalyst is Kimi K3, the 2.8-trillion-parameter open-weight model released July 27, which per Artificial Analysis delivers first-tier performance at roughly half the per-task cost of GPT-5.6 Sol on BrowseComp.
The terms are aggressive. Investors must fund by August 15, and direct entry to the cap table requires managing at least $500 million in assets; at a typical ~10% dilution, the raise would land near $6 billion. The company denied reports that it plans to file for a Hong Kong IPO this month, saying the G round is still underway and it could list within six months of closing. The fundamentals, from its prospectus: 2025 revenue of ¥1.699 billion with ¥278 million net profit, and 2026 H1 revenue projected at ¥1.052–1.128 billion, up 35.6–45.4% year over year. API revenue already exceeds 70% of the total.
The market context matters. Zhipu and MiniMax are already listed in Hong Kong, at roughly HK$437 billion and HK$71 billion market caps respectively, so a $50 billion Moonshot would top the sector. The interesting risk here is concentration: Moonshot's valuation is riding almost entirely on one model release, and the window between a model's buzz and its revenue proof is exactly where Chinese AI valuations tend to get volatile. If the round closes at $50 billion, it will be the strongest private-market signal yet that the K3 momentum is real money, not just hype.
Top comments (0)