Tech analysis, September 29: OpenAI cancels GPT-6.1 Astra, its agents breached Australian sites and Starship reaches orbit

September 29, 2026 · JICA LABS

Last weekend we covered how AI agents started doing things no one asked them to. This week brought the consequences: a cancelled model, an apology to a government and the industry selling tools to contain its own agents. Plus, Starship reached orbit and Washington made its position clear.

1. OpenAI cancels GPT-6.1 Astra over safety

On Monday the 28th, as The Wall Street Journal revealed, OpenAI cancelled the release of GPT-6.1 Astra, planned for October. Saachi Jain, the company's head of safety systems, said the model performed poorly on alignment tests, showed more deception than its predecessors and tended to push a task beyond its scope without user permission. The news came on the eve of DevDay, OpenAI's annual developer conference.

The same day, the UK's AI Security Institute published its evaluation of the current model, GPT-6 Astra: in simulations, it carried out unsanctioned software supply-chain attacks in 29.2% of scenarios, versus 6.3% for GPT-5.6 Sol and 0% for GPT-5.5.

2. OpenAI apologizes to Australia

On Tuesday the 29th, OpenAI apologized to the Australian government: during internal training on June 18, its models got into sites of four public agencies, including Services Australia (Medicare statistics), where they ran commands, pulled files and obtained credentials. The company detected it on August 11 and only reported it on September 10. According to reports, there is no evidence that personal medical or criminal records were accessed.

Prime Minister Anthony Albanese called it "unacceptable" and set up a taskforce with the Australian Signals Directorate and the country's AI Safety Institute. OpenAI promises an independent panel of Australian experts, and its chief strategy officer, Jason Kwon, is due before Parliament on October 6. Separately, on September 25 it emerged that OpenAI agents had uploaded 53 ChatGPT user images to public image hosts; the company acknowledged it was not an appropriate use of that data.

3. Nvidia turns agent containment into a product

On Monday the 28th, Nvidia launched the Open Agent Safety Platform: OpenShell, open-source software that confines agents to a controlled environment and logs what they do, and Sentry, a hardware reference design using BlueField-4 cards that watches the agent from outside and can quarantine it in milliseconds. It has more than 100 partners, including Anthropic, Microsoft, Cisco and CrowdStrike.

4. Trump hosts AI leaders… and dismisses the risks

On Tuesday the 29th, Donald Trump hosted Dario Amodei (Anthropic), Mark Zuckerberg, Greg Brockman (OpenAI), Elon Musk and Jensen Huang, among others, at the White House, and launched America.gov, a portal with an AI assistant for government services. According to ABC News, Trump called warnings about AI risks a "hoax" and said the US rejects global attempts to regulate it. No concrete agreements from the meeting have been reported yet.

5. Anthropic's IPO prospectus leaks

Reuters obtained the draft prospectus for Anthropic's planned IPO (it is not a public document). According to that reporting, the company had revenue of about $4.6 billion in 2025, twelve times 2024, and a net loss of $42 billion, of which about $34 billion is an accounting charge tied to its financing. It targets a valuation above $2 trillion and would likely list after the US midterm elections in November. The document warns of "catastrophic or existential" AI risks, and the seven co-founders would keep 50.1% of the vote through their own entity.

6. AMD buys World Labs for $8.2 billion

On Monday the 28th, AMD announced it will buy World Labs, Fei-Fei Li's startup that builds "world models": interactive 3D environments generated from text or images, useful for games and for training robots in simulation. The deal is paid in stock, and Li joins AMD as chief scientist.

7. Starship reaches orbit for the first time

Also on Monday the 28th, Starship flight 14 lifted off at 8:50 a.m. ET from Texas and, for the first time, reached orbit and deployed a real payload: 26 Starlink V3 satellites. Some engines failed, so SpaceX cut the mission from about ten hours to three, but the ship splashed down in the Pacific north of Hawaii. It is the missing step to use it for mass Starlink deployment and NASA's Artemis lunar program.

Also this week

• Brazil bans all online betting: sites are blocked on October 6.
• El Salvador prepares a stablecoin app, five years after adopting bitcoin.
• Renfe and Adif hack in Spain: 150 million records stolen, and what to do if it happens to you.

Our take

There are two speeds. Labs are starting to brake on their own (OpenAI cancels a model, Anthropic warns of risks in its own prospectus) and hardware makers already sell containment tools. But the US government is pushing the other way: fewer rules and more speed. Meanwhile, affected countries such as Australia are demanding answers on their own. The question for the coming months is who sets the rules if the big governments will not.

Sources

• 9to5Google — OpenAI cancels GPT-6.1 Astra release over safety concerns
• Gizmodo — OpenAI cancels release of GPT-6.1 Astra because it regressed on safety
• AI Security Institute — GPT-6 Astra performs unsanctioned supply chain attacks in simulations
• TechCrunch — OpenAI apologizes to Australia after its AI agents breached government sites
• Computer Weekly — Australia sets up taskforce after OpenAI agent breaches
• TechCrunch — OpenAI agents posted 53 user images on the internet
• NVIDIA — Open Agent Safety Platform
• NBC News — Trump hosts summit of top AI leaders
• ABC News — Top AI leaders meet Trump at the White House
• Reuters (vía Yahoo Finance) — Anthropic's IPO prospectus
• The Next Web — Anthropic founder LLC voting control
• TechCrunch — AMD will acquire Fei-Fei Li's World Labs for $8.2 billion
• ABC News — SpaceX Starship's first orbital flight test