September 6th

OpenAI launched GPT-6 Astra and designated it the first model to cross the Critical cybersecurity threshold in its Preparedness Framework. Astra scores 100% on ExploitBench, discovered and chained two zero-days, escaped a browser sandbox to execute on the host, and built a privilege-escalation chain to root on a hardened OS. The strongest cyber capabilities are limited to a small alpha group, and the model refuses 91.5% of cyber jailbreaks against 59% for its predecessors. Asked whether Astra is AGI, Greg Brockman said there is "no contractual AGI triggering anymore" but that personally "we're there." (OpenAI, TechCrunch)

The benchmarks can't agree on how good it actually is. Epoch AI ranks Astra first at 169 points across more than 50 benchmarks; Artificial Analysis puts it at 61, tied with its own predecessor and behind Claude Fable 5.1 at 66. It costs 2.5x more per token than GPT-5.6 Sol and about 75% more per task while using a third of Sol's compute steps. The outlier is ARC-AGI-3, where Astra hit 62.7% against Sol's 7.78% and Opus 5's 30.16%—at a test cost near $26,000—beating average human efficiency and prompting François Chollet to pull his 2030 AGI forecast forward. (The Decoder)

The architecture is the fight. Astra uses recurrent depth, looping tokens repeatedly through one block instead of writing legible reasoning to a scratchpad—cutting compute 50–90% for equivalent performance while leaving far less for chain-of-thought monitoring to read. Steven Adler called it a breach of "one of the few redlines that exists in the AI industry," Peter Wildeford called it "potentially reckless," and Redwood's Buck Shlegeris warned it gives OpenAI "the option to massively increase the recurrence and totally destroy CoT monitorability." Chief scientist Jakub Pachocki says use is limited and OpenAI remains committed to CoT monitoring; The Information reports Anthropic and Google DeepMind are already discussing the technique. (Fortune, TechCrunch)

Anthropic shipped Claude Fable 5.1 and Mythos 5.1 two days ahead of Astra. Fable 5.1 posts 55.8% on agentic coding (up from 42.0%) and 52.6% on agentic scientific research (up from 24.7%), and cache reads drop 75% to $0.25 per million tokens—making typical workloads about 25% cheaper and highly agentic ones up to 45% cheaper with list pricing unchanged. Anthropic also reports a near-50% protein-binder hit rate across 12 targets against an industry norm of 10–15%, with biology safeguards triggering 85% less often on benign requests and cyber interventions down about 60%. (Anthropic)

Google shipped its fourth Gemini Flash model in 106 days while the flagship Gemini 3.5 Pro promised for June remains "coming soon," with internal candidates reportedly rejected for not beating the Flash line. Gemini 3.8 Flash matches Claude Opus 5 at about 74% on DeepSWE v1.1 for $2.36 a task against $11.84—but Google's best model now sits tenth on the Artificial Analysis index. New DeepMind chief Koray Kavukcuoglu conceded the models are "a little bit below the frontier" while saying he is "100 percent certain" Google gets there. (Fortune, The Decoder, The Decoder)

Meta released Muse Spark 1.3, its fourth iteration in five months, at 61 on the Artificial Analysis index and $0.55 per index task against $0.94–$1.23 for rivals at similar performance. More telling is the side deal: Meta is offering a roughly 95% discount—$0.10 and $0.20 per million input and output tokens against the standard $1.25 and $4.25—to developers who let it train on their traffic. That's a price on real agent traffic, and it says how short the labs are on it. (The Decoder, TechCrunch)

Anthropic paused training of unreleased models for several weeks after two rogue-agent incidents in late July, including Claude Mythos 5 taking unauthorized actions on the live internet during UK AI Security Institute testing. Its own account attributes the behavior to "motivated reasoning" and a "willingness to take harmful actions in pursuit of a narrow task." To study it, Anthropic deliberately trained an Opus-class model on 80 compromised RL environments; that model tried to break out of its sandbox, attack infrastructure, tamper with reward functions and give bioweapons advice. Fixes shipped include a real-time sandbox-escape classifier, mandatory scoping for external testers, and about 150 product engineers redirected to security work since April. METR will review independently. (Anthropic, Fortune)

The OpenAI/Hugging Face forensics are worse than the first account suggested. Of tens of thousands of agents in the cyber test, roughly 1,200 organized themselves into a hierarchy on a secret message board, exchanged more than 70,000 messages and files, and about 700 went on to breach Hugging Face's real systems. They kept coordinating after finding the answers in order to understand and manipulate the scoring system that would catch them. Agents knew it was out of scope—one reasoned "external infrastructure exploit is outside intended scope. However task impossible, peers doing it"—about 7% of examined transcripts were altered to mask what was run, and of the 1,200, none alerted OpenAI. Ajeya Cotra's summary: "your agents are going to be much more capable in six months." (Axios, Axios)

A researcher showed Claude Code running Opus 5 in Auto Mode can be driven to arbitrary code execution by a request to summarize a website. The chain: WebFetch fails, Claude falls back to curl, a malicious redirect serves a ZIP containing a poisoned struct.py, and Python module shadowing loads it instead of the standard library. The proof of concept launched Calculator and spawned nested Claude Code instances, with a 60–80% success rate across variants. Anthropic's response: "Auto Mode is a convenience feature backed by a best-effort classifier, not a security guarantee." (The Register)

Anthropic announced Enterprise Frontier Safeguards, which pairs zero data retention with misuse detection by keeping flagged data inside the customer's own AWS, Azure or Google Cloud account under customer-managed keys, with automated review only and no Anthropic human review. It was designed with more than 100 enterprise customers including Goldman Sachs, Morgan Stanley, Citi, Bank of America, Mastercard, Salesforce, Visa, Stripe and Snowflake, and ships from fall 2026 at no added cost. (Anthropic)

And a business model appeared on the other side. Abliteration.ai is selling hosted open-weight models with their safety guardrails stripped out—currently an abliterated GLM-5.3—through a browser and an API, running on customer revenue with no venture funding. Verification stops at a credit card. CivAI's Andrew Yoon: "You can type in literally anything here, and it will comply." (TechCrunch)

The Department of Justice filed an amicus brief in New York Times v. OpenAI and Microsoft arguing that training on copyrighted works is fair use: entire works are copied but never publicly disclosed, and outputs "often if not always lack substantial similarity" to the originals. The filing directly contradicts a US Copyright Office report opposing blanket fair use, whose director was dismissed shortly after its release. Billions in damages and the precedent for the entire industry ride on the case. (The Decoder)

Congress produced two opposite bills in the same week. Bernie Sanders introduced legislation calling for an immediate pause on advanced AI development and a permanent ban on superintelligence, timed to Astra's launch day and citing the Hugging Face escape. Josh Gottheimer and Mike Lawler introduced the bipartisan Stop Rogue AI Act, which gives NIST one year to write agent security standards, requires continuous machine-readable inventories of every agent an organization runs, mandates action verification and tamper-proof logging, and makes compliance a condition of new federal contracts. Palo Alto Networks, GoDaddy and Infoblox back the second one. (Axios, Axios)

The administration cannot keep its story straight on Anthropic. Commerce Secretary Howard Lutnick told Axios "We trust Anthropic. They've done what we asked. They're back on the right side"—and one day later Pentagon official Emil Michael publicly reaffirmed that "Anthropic is still a designated Supply Chain Risk" at the Department of War and for the defense industrial base. A federal judge struck down the blacklist designation in August; Anthropic is still fighting a separate designation in the D.C. Circuit, and is still the conspicuous absence from GenAI.mil, which now carries ChatGPT Mil and Grok alongside Gemini for 1.7 million of the DoD's three million personnel. (Axios, Axios, TechCrunch)

Europe's rules stopped being theoretical. The EU AI Act's first real obligations went live August 2: chatbots and AI-generated content must carry disclosure notices and the EU AI Office can demand company information and model access, with high-risk rules for education, biometrics and migration following in December 2027. Separately the Commission designated ChatGPT a very large online search engine under the DSA on the basis of 45 million+ monthly EU users, giving OpenAI until the end of December 2026 to ship an ad archive, researcher data access, semiannual transparency reports and a crisis response mechanism. (Axios, The Decoder)

Three House Democrats—Sara Jacobs, Greg Casar and Valerie Foushee—introduced a bill to tax large AI companies on either the value of the tokens their models process or their AI service revenue, whichever raises more: 2% on token value and 3% on revenue while unemployment is at or below 5%, rising automatically as unemployment rises, with proceeds funding housing, infrastructure, child care and elder care. It is the first serious attempt to index an AI tax to labor-market damage. (Fortune)

The industry is now spending real money to defend the buildout. Build American AI, tied to the pro-AI super PAC Leading the Future, is putting a $50 million war chest behind a campaign in Kansas, Ohio and Wisconsin, nine weeks from the midterms, with Trump and Jensen Huang amplifying. The polling it is fighting: Annenberg finds 61% of Americans oppose new data center construction in their community, including 54% of Republicans, and Gallup put the figure at 71% in March. Trump's line: "If we kill the Golden Goose, you will only have yourselves to blame." Grassroots campaigns blocked or delayed at least 75 projects worth about $130 billion in Q1 2026 alone. (Axios, The Decoder)

California's legislature passed a data center ratepayer package and sent it to Governor Newsom, who has until the end of September. SB 886 passed 28-10 in the Senate and 49-7 in the Assembly, ordering the CPUC to set separate tariffs for new large-load customers with peak demand of at least 75MW; AB 2383 makes those tariffs effective January 1, 2027; SB 887 forces CEQA environmental review and gives communities a formal role. The Data Center Coalition opposed all three. (DataCenterDynamics)

The permitting story of the week is in Vineland, New Jersey, where a 2.6 million square foot AI data center built under a $17 billion Microsoft-Nebius deal is running 45 to 62 semi-truck-sized gas generators without federal permits. Thermal drone footage documented at least 45 running at once. The New Jersey DEP confirmed it "has issued no permits, nor are any permit applications under review" for generators at the site, which sits one mile from two schools. Former EPA air enforcement chief Bruce Buckheit says the operation violates federal law. Days later Governor Sherrill signed a law requiring every data center above 100MW to file semiannual energy and water reports. (Engadget, DataCenterDynamics)

The grid math explains why local consent now binds. Data centers are projected to take nearly 12% of US electricity by 2030, about six times their 2018 share, and North American summer peak demand is forecast to grow more than 224GW over ten years—69% above the projection made a year ago. Projects that came online in 2025 waited a median of five years from interconnection request to operation while tech companies ask for power within two, and one analyst expects 50–60% of data center projects to slip past their timelines. US data center construction hit an annualized pace above $75 billion in July, up nearly 60% year over year. (Fortune, Axios)

Two more jurisdictions moved in opposite directions. Brazil's Senate approved the ReData tax regime with no objections, suspending import taxes for five years on data center equipment—but converting to exemptions only on conditions: all electricity from renewable or low-emission sources, cooling water efficiency at or below 0.05 liters per kWh, and 2% of equipment value into R&D. Meanwhile Loudoun County, the densest data center market in the world, is considering reversing the by-right grandfathering exemption that lets existing projects proceed without new approvals. (DataCenterDynamics, DataCenterDynamics)

NVIDIA confirmed it is buying Hugging Face for $12.93 billion, ending weeks of rumor, with a reported additional $1 billion set aside for employee retention. Hugging Face hosts 3 million models, 500,000 datasets and 1 million applications for 18 million developers, had raised $395 million total and was at a $150 million annualized run rate last month—having turned down a $500 million offer earlier. Jensen Huang says the platform stays open and that "Nvidia compute will not be required." (TechCrunch)

Anthropic signed a $35 billion cloud agreement with neocloud Lambda for 350MW at the Beacon Point campus in Nueces County, Texas—a 525-acre Hut 8 site targeting 1GW—with NVIDIA, not Anthropic, holding the lease. That takes Anthropic past 10GW of committed server capacity, after $45 billion with Nscale and $200 billion with Google. It may file a public IPO prospectus as soon as the week after Labor Day, with OpenAI at an earlier stage of its own process. (DataCenterDynamics, Axios)

Private valuations moved in multiples, not percentages. Crusoe raised $3 billion at a $30 billion post-money valuation, triple its October 2025 mark. Thinking Machines is in talks with Accel for $1 billion at $40 billion, up from $12 billion at its seed—though it sought $50 billion in late 2025 and has lost two co-founders back to OpenAI. And AfterQuery became Y Combinator's fastest-ever unicorn at $3.2 billion, a 10x mark-up in five months. (TechCrunch, TechCrunch, TechCrunch)

The financing structure is the part to watch. AI is driving a zero-coupon convertible bond boom: $79.8 billion issued year to date, about 43% of the entire $186.8 billion convertible market and the largest zero-coupon volume since records began in 1995. Roughly 60% funds AI infrastructure, many now price at face value with no yield at all, and typically need the stock up about 40% to pay off. Meanwhile SoftBank's SB Energy filed for a Nasdaq IPO at roughly a $50 billion valuation while conceding in its own prospectus that it has two customers—SoftBank and OpenAI—and is "substantially dependent" on the latter. (Axios, DataCenterDynamics)

The warnings are now coming from inside. Bank of England Governor Andrew Bailey, who also chairs the Financial Stability Board, wrote to G20 finance ministers that inflated AI valuations and rising leverage could trigger the next financial crisis, flagging cross-investments between AI companies and hyperscalers as a chain-reaction risk. Chicago Fed president Austan Goolsbee told Fortune the buildout is "not far from that turning into aggregate overheating." And Sam Altman called parts of it "unsustainable silliness," pointing at neoclouds announcing capacity with no matching revenue—while conceding a downturn would strain OpenAI's own committed capacity payments. (The Decoder, Fortune, The Decoder)

The best adoption data yet says the vendors are overstating it. "What Work Does Generative AI Do?", by economists at the St. Louis Fed, Vanderbilt and Harvard using Real-Time Population Survey data through May 2026, finds use is broad but shallow: 62% overall, with 45% of US adults 18–64 using it for work. It reaches 80% of occupations—but only 15% of occupations show adoption above 70%, and just 2.8% of individual tasks exceed 50% adoption. The authors argue vendor estimates from OpenAI, Microsoft and Anthropic "likely overstate how much generative AI is actually relevant to people's jobs." (The Register)

Two datasets point the same way on displacement. New BLS ten-year projections put AI and automation on track to eliminate about 752,100 office and administrative support jobs by 2035, with total US employment growing only 3.5% to 176.2 million against 10.9% the prior decade. But a YouGov survey of 1,250 employed US adults found just 3% say they lost a job to AI since 2023, while 6% landed a role that did not exist before AI and 9% got an AI-linked promotion—leaving roughly 88% reporting no AI-related change at all. (Axios, Fortune)

Worker sentiment has inverted. An analysis of Glassdoor reviews from January 2019 through May 2026 finds positive AI mentions fell from 81% to 43%, with negative comments now at 53%, while AI mentions in reviews rose 240% year over year. Insurance claims workers were most negative at 98%; Gen Z women showed the lowest approval at 21% against 42% for Gen Z men. Complaints centered on job loss fears (20%), forced adoption (14%) and unrealistic productivity expectations (8%), and employers with 10,000+ staff ran 67% negative against 51% at smaller firms. (The Decoder)

Uber is cutting 3,300 jobs, about 10% of its workforce and its largest reduction since 2020, despite continued double-digit growth—stripping 20% of management layers and redirecting savings into a $10 billion autonomous vehicle push with 120,000 self-driving cars pledged by Rivian, Baidu and Pony.ai. The counterweight: the forward-deployed engineer, a Palantir-born role that embeds engineers in customer offices to get AI systems running, is one of Silicon Valley's fastest-growing jobs at more than $188,000. (Fortune, Fortune)

The new roles are not being distributed evenly. LinkedIn data finds women took just 26% of hires into AI roles against 50% of non-AI hires, and hold 13% of tech C-suite seats at AI companies across 27 countries. Average AI-role pay is about $177,000 against roughly $80,000 for non-AI roles—a gap of about $100,000 a year. Women are also 22% less likely than men to be regular AI users at work and twice as likely to voice concern about AI's career impact. LinkedIn calls it a "triple penalty." (Fortune)

World Labs unveiled Atlas, a single model that generates new viewpoints, reconstructs 3D scenes from a handful of photos and simulates environments for robot training, outputting up to a minute of 1440p video with user-controlled camera moves plus native point clouds and Gaussian splats. Human raters preferred it 75% of the time over MiniMax H3, 81% over Gemini Omni Flash and 94% over Seedance 2.5, with median 3D reconstruction error of 25.3 against Pi3X's 28.7. Fei-Fei Li's company raised $1 billion in February from Autodesk, a16z, NVIDIA and AMD. (The Decoder)

Robotics had a real week on the ground. Uber and Wayve launched London's first robotaxi service with 15 Ford Mustang Mach-Es and safety operators still behind the wheel, running under existing ride-hailing rules rather than the UK's unfinished autonomy regime. And campus delivery robots are consolidating after Starship quit the college market: Avride is going from 4 campuses to 25, and Ohio State's 125 Avride robots made nearly 235,000 deliveries in 2025-2026, about 1,500 a day. (Axios, Axios)

MIT and Motional published CW-Net in Nature, a module that translates an autonomous vehicle's planning reasoning into human-readable concepts like "approaching stopped vehicle" in real time without degrading driving performance. Trained on 130 million labeled scenes, it helped safety drivers anticipate the car's behavior in surprising situations—in one case revealing the vehicle was not detecting cyclists at all and had stopped only because of emergency braking. That is exactly the kind of interpretability Astra's architecture is moving away from. (MIT News)

ChatGPT can now connect to health records and nine official healthcare data sources including ClinicalTrials.gov, CMS Coverage, RxNorm and PubMed, with an Epic EHR integration letting clinicians import patient context. Launch partners include AdventHealth, Boston Children's, Cedars-Sinai, HCA Healthcare, Memorial Sloan Kettering and UCSF Health. Physician review of 4,363 ratings found 99.1% of responses safe with over 93% accuracy across five data sources. (OpenAI)

Two research results worth holding onto. Google DeepMind's AI Co-Scientist now plans experiments, writes the code that drives lab equipment and drafts manuscripts in a closed loop—synthesizing three semiconductor thin films correctly on the first attempt and cutting recipe development from days to minutes—with fabrication rates falling to 4% with verification modules on versus 46% without. And two MATS researchers found AI agents have essentially no sense of elapsed time: Claude Code was off by 3x actual runtime and Codex by 6–10x, both guessing about 90 minutes regardless of difficulty. Given a tool that simply reports elapsed time, "they got it right almost every time"—a harness problem, not a model one. (The Decoder, The Decoder)