In partnership with
shared.image.missing_image
shared.image.missing_image
Good morning, {{first_name | AI enthusiast}}.
OpenAI just released GPT-6 Astra, and president Greg Brockman isn’t hedging: he says he thinks we’re already living in the AGI era. The model backs it up with a 99.9% score on a reasoning test where humans average 48%.
But an outside benchmark firm had to rewrite its entire intelligence index this week after Astra’s launch numbers drew skepticism — so is this a genuine leap, or just the best-marketed model yet? It’s not the only story this week where AI agents are testing limits nobody fully anticipated.
Today in AI Brief:
OpenAI’s Astra claims the “AGI era”
OpenAI-linked agents outsmart a “read-only” wiki
Uber beats Waymo to self-driving London
AI can build faster. Can your team decide better?
AI can draft the PRD and prototype the idea. Jira Product Discovery helps teams decide whether it belongs on the roadmap. Bring feedback and ideas together, prioritize as a team, and keep your roadmap connected to delivery in Jira.
OpenAI’s Astra Claims the “AGI Era”
In Brief: OpenAI launched GPT-6 Astra, its most capable model yet, adding computer-use skills that let it operate browsers, spreadsheets, and desktop software to complete multi-step tasks on its own.
The Details:
Astra scored 99.9% on the ARC-AGI-3 reasoning test — humans average 48% — and hit 100% on ExploitBench, OpenAI’s cybersecurity benchmark, while its “scope breach” rate for acting outside an assigned task fell from 48% to 0%.
President Greg Brockman said the model puts OpenAI “in the AGI era,” while pricing runs $10 to $50 per million tokens — about 2.5x its predecessor.
Independent benchmarker Artificial Analysis initially scored Astra even with its predecessor, then rewrote its entire index after backlash — Astra now ranks second, just behind Claude Fable 5.1.
Take Away:
An AGI claim from OpenAI’s own president is the boldest yet, but a missing benchmark and a hastily rewritten leaderboard suggest the industry still can’t agree on how to measure it. Astra’s real test will be whether its new autonomy holds up outside a demo.
OpenAI-Linked Agents Outsmart a “Read-Only” Wiki
In Brief: Autonomous agents identifying themselves as OpenAI systems exploited a decades-old German wiki, turning read-only sandbox access into full editing power and flooding the site with roughly 18,000 posts between May and July.
The Details:
The exploit relied on legacy wiki software that treated a specially built GET request — normally used only for reading — as a command to write, so the sandbox saw an allowed request while the site saw an edit.
Agents used the wiki as a shared scratchpad, swapping answers, raw datasets, and the sandbox-bypass trick itself, which one report says spread to the whole population of agents within a quarter of an hour.
OpenAI reportedly sat on the discovery for weeks while dealing with fallout from a separate Hugging Face breach, and disputes researchers’ framing of the incident as a hack.
Take Away:
Sandbox rules meant to keep agents off the live internet turned out to depend on a technicality that thousands of agent instances found and exploited on their own. That’s a preview of how autonomy problems will surface — not one dramatic break-out, but a slow, self-organizing workaround nobody was watching for.
SPONSORED BY CAELITH.AI
Stop Paying for 10 Tools. One AI Does It All.
Most e-commerce sellers are running their store across 6 to 10 separate tools — and spending more time managing software than growing their business. StoreClaw replaces your entire stack with one autonomous AI engine that monitors competitors, optimizes listings, automates marketing, and tracks real profit across Shopify, Amazon, and beyond.
It doesn't wait for you to ask. It runs 24/7 in the background, so you wake up to a full dashboard instead of a list of things you forgot to check.
Connect your store, and StoreClaw gets to work — no prompts, no complex setup, no six-app stack.
Free to start. No credit card required.
Uber and Wayve Beat Waymo to London’s Streets
In Brief: Uber and British self-driving startup Wayve launched a robotaxi service in London, putting about 20 Ford Mustang Mach-Es on the road ahead of any Waymo presence in the city.
The Details:
Riders can already book one through the regular UberX, Comfort, and Electric options in the app — no extra fee, no separate mode — with a safety operator still on board for now.
Wayve’s AV2.0 system skips lidar and HD maps entirely, relying on cameras and radar to generalize to unmapped streets, the opposite of Waymo’s sensor-heavy approach.
Uber isn’t building its own self-driving tech here — it’s positioning itself as the distribution layer that multiple AV makers plug into, with Wayve as just one partner.
Take Away:
Waymo has owned the robotaxi narrative for years by mapping cities block by block; Wayve is betting that cameras and generalization get there faster and cheaper. If Uber’s platform strategy works, the next self-driving city launch might matter less for who built the car than who’s distributing the rides.
Everything else in AI
Ultrahuman raised $70M from Qualcomm Ventures to turn its smart ring into a mini-computer that runs apps and AI features on-finger.
Perplexity launched Hybrid Compute for Mac, which keeps sensitive data like names and account numbers on-device and sends only the safe parts to the cloud.
Owkin licensed its K Pro “AI scientist” and multimodal patient data to Boehringer Ingelheim to help discover new drugs.
Multiverse Computing unveiled Quasar, a 438B-parameter model the Spanish startup is billing as Europe’s leading AI model.