Three AI Chatbots Crashed at Once, the Same Day OpenAI Launched One Built to Hack

On Thursday, three of the biggest AI chatbots on the planet stopped working within roughly the same half hour, and none of the companies behind them fully explained why. Hours earlier, OpenAI had released a new model built specifically to break into computer systems, restricted to vetted researchers because it is powerful enough to worry its own maker. Halfway around the world, India had a far simpler kind of good day: a rocket lifted off exactly on schedule and put a satellite into orbit that will watch the country almost continuously for years.

Three unrelated AI companies broke at the same time, and nobody has fully explained why

Error rates for ChatGPT, Claude and Grok all started climbing within the same half hour on the morning of September 3. By mid morning, Downdetector had logged more than 35,000 US reports for ChatGPT, plus well over a thousand each for Claude and Grok.

OpenAI said ChatGPT and its coding tool Codex were hit with elevated error rates. Anthropic said two of its heaviest models, Opus 4.8 and Opus 5, kept failing even after its other Claude models had recovered, and did not mark the incident resolved until around 12:16pm Eastern time. xAI confirmed Grok was down too and said it was working on a fix.

Nobody landed on one shared cause. Microsoft Azure reported a separate networking problem in its East US region during the same window, and some outlets pointed to that as a likely trigger, since a lot of AI infrastructure runs on the same handful of cloud backbones. OpenAI, Anthropic and xAI never confirmed a common root cause. What the incident did make plain is how much of the internet’s daily workflow now runs through a small number of AI providers sitting on a small number of cloud platforms. When one link wobbles, three unrelated companies’ products can go dark within the same half hour.

The same day, OpenAI released a model built to hunt for exploits

Hours before the outage, OpenAI began rolling out GPT-6 Astra, which it calls the most capable model it has ever shipped. The rollout itself was messy: reporters at CNBC and other outlets published details of Astra before OpenAI’s own page describing it was public, so the news broke ahead of the company’s own schedule.

What sets Astra apart is not a benchmark score so much as a warning label. OpenAI classified it as reaching “Critical” under its Preparedness Framework for cybersecurity, the first time any of its models has hit that tier. In testing, Astra found two previously unknown software flaws, the kind researchers call zero days because nobody has patched them yet, and chained them into a working exploit largely on its own. On ExploitBench, a benchmark built to measure whether a model can find and use real vulnerabilities, Astra scored 100 percent.

Because of that, OpenAI is not handing Astra’s sharpest capabilities to just anyone. Its most advanced cyber features sit behind Daybreak, an invitation only program for vetted security teams, with Cisco, Cloudflare and Palo Alto Networks named as early partners, and a hardware security key required on every individual account. Broader access through ChatGPT’s paid tiers, the API and Amazon Web Services is due in the following days, without the exploit building features attached. OpenAI is charging 10 dollars per million input tokens and 50 dollars per million output tokens, with a faster processing mode available at double that price again.

OpenAI president Greg Brockman called Astra the company’s most intelligent model yet, pointing to gains in coding, computer use and scientific reasoning alongside the jump in cyber capability. Whether a model that can find working exploits faster than most human researchers ends up helping defenders more than attackers is the bet the entire Daybreak program is really built on.

India put its first geosynchronous eye in the sky

Away from the AI industry, ISRO had a clean morning. At 2:55am on September 4, a GSLV-F17 rocket lifted off from the Satish Dhawan Space Centre in Sriharikota carrying EOS-05, a satellite weighing 2,367 kilograms, the heaviest Earth observation spacecraft India has built. Nineteen minutes later, EOS-05 separated into a sub geosynchronous transfer orbit, on its way to a working position roughly 36,000 kilometres up.

EOS-05 matters because of where it sits, not just what it carries. Most of India’s Earth observation satellites orbit much closer to the ground and pass over any given spot for only a few minutes at a time before moving on. From geosynchronous orbit, EOS-05 can watch the same stretch of the Indian subcontinent almost continuously, which ISRO says will help track agriculture, water resources, forestry, ocean conditions and urban growth closer to real time than before. It is India’s first satellite built to do that job from that altitude, after an earlier attempt, EOS-03, was lost in 2021 when its cryogenic engine stage failed to fire correctly.

Prime Minister Narendra Modi called the launch “yet another outstanding achievement by ISRO and a proud moment for our nation” in a post on X, describing it as a reflection of the growing partnership between ISRO and Indian industry in building out the country’s space capabilities.

Three companies whose products a big share of the internet now leans on broke down at the same time this week, for reasons none of them has fully explained. Nineteen minutes after a rocket lit up the sky over Sriharikota, a satellite that will watch India continuously for years reached orbit without a hitch. That is what infrastructure looks like right now: some of it fragile in ways nobody predicted, some of it working exactly as planned, often on the very same day.