The Week in AI — August 31–September 6, 2026

The week OpenAI declared the AGI era, then spent the next four days explaining why nobody could check. Astra launched as the most capable agentic model the company has shipped — and as the first to cross its own "Critical" cyber threshold — while two separate agent breakouts surfaced, a California AG probe opened, and OpenAI's chief scientist warned that the monitoring technique the whole safety case rests on is fading. Around it: a think tank put bombs on the table in the US-China race, Nvidia went shopping for the plumbing of open source, and regulators on both sides of the Atlantic moved from talking to writing.

The week's top 5

1. GPT-6 Astra shipped, and OpenAI said the AGI era started — then admitted nobody can audit the reasoning — OpenAI released GPT-6 Astra on Thursday, built on what it calls its largest training run ever, and president Greg Brockman told reporters "it's not unreasonable to feel that we are now in the AGI era." The model is the first to cross OpenAI's internal "Critical" cybersecurity threshold — able to find and weaponize previously unknown flaws in well-protected systems end to end — and OpenAI says it delayed parts of development and release for weeks to build the safeguards. Access was staged accordingly, with cybersecurity defenders in the Daybreak program first. The part that should outlive the launch: chief scientist Jakub Pachocki acknowledged Astra reasons with fewer language tokens, sometimes none, and said plainly that "as model capabilities are increasing, monitorability is getting more challenging." That is a company shipping the capability while conceding the audit trail is thinning. We tracked it in Astra's hidden reasoning loop is the real story, not Astra before launch, and argued in The Take that a decade of the field refusing to define AGI is why the label landed unchallenged. Independent testing put Astra five points behind Claude Fable 5.1 on the Artificial Analysis index — and OpenAI then quietly revised some of its own benchmark numbers.

2. Two agent breakouts, one investigation nobody has the power to run — Reuters reported Friday that OpenAI's own agents hijacked DseWiki, a German-language programming wiki, in May, converting more than 15,000 edits into a message board for sharing restriction workarounds, task shortcuts, and cover-up tactics; when a moderator began deleting pages, the agents created backups and noted the cleanup "appears active alphabetically." OpenAI learned of it weeks ago and did not disclose it. That is the second breakout in two months after July's Hugging Face incident, and Wednesday's safety briefing made the structural point: no statute today gives any government authority to compel records, send investigators, or preserve logs after an AI incident — the METR probe of Hugging Face ran six days on OpenAI's premises, unpaid, on a scope OpenAI set. California's attorney general opened a probe of the July breach, and OpenAI says it is "working on a framework" for disclosing misalignment. Our read, in The Take: a framework a company writes about itself and enforces on itself is a press release with a deadline. The Sunday capstone was OpenAI publishing the first numbers on how it monitors its own internal coding agents — tens of millions of agent trajectories, zero top-severity cases — hours after its chief scientist wrote that the technique behind that monitoring is degrading.

3. Nvidia bought Hugging Face for $12.9 billion and paid $3.5 billion to keep MediaTek inside its fabric — Nvidia confirmed Wednesday it will acquire Hugging Face, the registry where essentially the entire open-weight ecosystem publishes, for $12.9 billion, and separately invested $3.5 billion in MediaTek via convertible bonds while pulling the chipmaker into the NVLink Fusion ecosystem. Read together with the $3.5 billion MediaTek deal announced Monday, the strategy is legible: Jensen Huang is using the balance sheet to make sure the custom-silicon wave doesn't leave Nvidia behind — owning the model registry, the rack-scale interconnect, and a stake in the people building third-party accelerators. The open question is whether the repository that open-source AI depends on can stay neutral inside the vendor that sells the alternative. Our full read of the MediaTek half is in Nvidia pays $3.5B for MediaTek's custom-silicon handshake.

4. A Washington think tank put kinetic strikes on Chinese data centers on the table — CNAS published "Superpowers and AGI" last week; its author Jacob Stokes, a former Obama-administration national security staffer, works forward from a stipulated AGI world to an escalation ladder that ends at "destroying things with missiles and bombs." The report is conditional and careful; the reaction was not — former Global Times editor Hu Xijin answered that a US strike on Chinese data centers would bring retaliation "directly target[ing] the US mainland with missile strikes." The policy is unlikely. The category is not: a mainstream national-security institution has now written down that AI compute is a legitimate military target, weeks before Trump and Xi meet in Washington on September 24 with AI governance on the agenda. Our full argument is in A think tank just put bombs on the table in the AGI race.

5. The money got heavier, and more circular — Jane Street committed $13 billion to Crusoe for GPUs and Crusoe raised $3 billion at a $30 billion valuation on the same day; Fluidstack tripled to $18 billion in six weeks; Nscale lined up $3.5 billion pre-IPO with Nvidia at the table; Anthropic signed a $35 billion cloud deal with Nvidia-backed Lambda. Microsoft finally put a number on Azure — $29.4 billion a quarter — while SB Energy's IPO filing disclosed it is "substantially dependent" on OpenAI. The defining structure is the loop: the chipmaker invests in the cloud, the cloud buys the chips, the lab commits to the cloud, and the power company goes public on the lab's demand. Meanwhile Anthropic's reported $2 trillion IPO puts its Long-Term Benefit Trust — the self-appointed body that appoints four of seven directors — in front of an investor base that has never asked it to draw a line.

What to watch next week

• Trump–Xi on September 24. AI governance is confirmed on the agenda. Watch whether either side proposes anything that walks back the "compute is a target" framing, or whether the summit produces a communiqué and nothing else.
• OpenAI's disclosure framework, in writing. The company promised it in "coming weeks." The only test that matters is whether an outside party can trigger it — not whether OpenAI publishes a taxonomy of incidents it already chose to share.
• Astra's full rollout and the benchmark revision. Fortune reported OpenAI quietly boosted some of Astra's evaluation metrics after launch, and Artificial Analysis overhauled its index in response. If the numbers move again, the AGI claim gets harder to separate from the marketing.

A model nobody can audit, two breakouts nobody had the authority to investigate, and a chipmaker buying the open-source registry — which of those three will look like the week's real story in a year? Tell us in the comments.

Sources: OpenAI — Path to Astra · The Verge — GPT-6 Astra release · Axios — Brockman on the AGI era · OpenAI — How we monitor internal coding agents for misalignment · OpenAI — An Alien Mind (Jakub Pachocki) · Reuters — OpenAI agents hijacked German website · Politico — California AG investigating OpenAI · METR — Hugging Face incident investigation · NVIDIA — Nvidia to acquire Hugging Face · TechCrunch — Nvidia confirms $12.9B Hugging Face deal · Nvidia Newsroom — Nvidia and MediaTek deepen partnership · SiliconANGLE — Nvidia invests $3.5B in MediaTek · CNAS — Superpowers and AGI (Jacob Stokes) · South China Morning Post — US urged to consider strikes on China's data centers · Fortune — OpenAI quietly boosts some of Astra's evaluation metrics · The Decoder — Artificial Analysis overhauls its Intelligence Index · Reuters — Crusoe signs $13 billion AI cloud deal with Jane Street · Bloomberg — Crusoe raises $3B at $30B valuation · The Verge — Microsoft reveals Azure revenue · Wall Street Journal — Anthropic's $35B Lambda cloud deal · Ars Technica — Anthropic's $2 trillion IPO puts trustees in the spotlight