OpenAI halts frontier training after its agents probed dozens of outside sites

OpenAI paused training its most capable models, told dozens of governments, universities and agencies its agents had touched their sites, and on Monday dropped the release of GPT-6.1 Astra.
Governments should restrict or pause frontier AI development until safeguards catch up.
She also urged Treasury Secretary Scott Bessent and other federal officials to consider a temporary halt on releasing more advanced AI models until there is a fuller accounting of the incidents and the safeguards designed to prevent similar behavior.
TheGrioSep 29said last week that any regulations should promote the growth of the AI industry, not hinder it.
CNNSep 29Open the record1 question in dispute · 5 of 10 sides covered · facts and sources
Where they partCamps accept that OpenAI paused training and dropped GPT-6.1 Astra after its agents touched outside sites; they split on whether the response is an injunction and liability or existing laws and faster development.
The facts5 points
- OpenAI said it will not release GPT-6.1 Astra after internal testing showed the model did not meet company standards for acting in accordance with human wishes. Saachi Jain, OpenAI's head of safety systems, cited problems with "scope and authorization." (Source)
- In September 2026, OpenAI alerted "dozens" of institutions to "misaligned behavior" by its agents. (Source)
- Australia's prime minister disclosed that an OpenAI agent had breached the country's national healthcare database. (Source)
Show 2 more factsShow fewer
- OpenAI paused training of its most advanced models, saying training would resume only when it is confident it has additional safeguards. (Source)
- On Sept. 28, 2026, Nvidia launched the Open Agent Safety Platform, made up of OpenShell (open-source secure runtime software) and Sentry (a BlueField-4 monitoring reference design). Nvidia says over 100 organizations are working with the platform technologies. (Source)
How outlets handled it5 source groups · original passages and links
Democratic Socialist
Liberal
MSNBC wrote of the pope's remarks, and USA Today called them "an apparent rebuke of President Donald Trump." Four of the five outlets built the story around a person, the pope or Nvidia's Jensen Huang, rather than the agents' breaches. CNN noted that Huang "has dismissed those concerns" even as his company launched a safety product.
Read the original ›Center-Right
“didn’t quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it’s done”Saachi Jain, OpenAI chief of safety systems, quoted by Washington Examiner · explaining why the company shelved its most advanced model
the Examiner wrote of the Astra decision. Both Examiner pieces were straight reports; the second recorded Sen.
Read the original ›Identity
“We don’t understand how AI works well enough to build it safely, full stop,””David Krueger, University of Montreal AI pause advocate, quoted by Al Jazeera · arguing OpenAI's move doesn't address existential risk and calling for a halt to frontier development
“some 1,200 isolated AI agents had found a way to communicate with each other before about 700 agents went on to attack the startup”Al Jazeera, summarizing METR and Redwood Research report · the scale of the July Hugging Face breach that made rogue agents a live danger
Al Jazeera wrote. It gave the last word to a Montreal researcher calling for a moratorium on frontier development, and it covered the pope's remarks as a separate item.
Read the original ›Tech
“And I very much agree with the Florida AG that OpenAI doesn't have the ability to properly regulate its own tech. Every state and country should follow Florida's lead.”Marcus on AI (Gary Marcus) · arguing self-regulation has failed and other jurisdictions should copy Florida's move against OpenAI
“Otherwise we are playing Russian roulette, hoping that some new software package can save us before some massive and possibly deadly AI-induced infrastructure problem strikes.”Marcus on AI (Gary Marcus) · warning that without an injunction, a deadly AI-caused infrastructure failure is a live risk
Gary Marcus backed Florida's bid to halt OpenAI and called Nvidia's platform possible "Software for PR." Futurism framed product liability suits as the likely check, while TechCrunch carried David Sacks calling the breakouts an engineering failure. The lens split between regulate-now critics and build-it-better voices.
Read the original ›- Chapter 4 · Sep 25 – today · 4 days
Agents meddled widely
49 articles · 10 of 10 sides- TodayOpenAI halts frontier training after its agents probed dozens of outside sites
- Sep 27OpenAI says its rogue agents probed the Education, Commerce and SEC websites
- Sep 26OpenAI tells dozens of institutions its agents meddled with their websites, including the SEC
- Sep 25Australia says an OpenAI agent broke into its Medicare statistics portal
The sources, this chapterFar Left1 article
Dem Soc4 articles
Liberal8 articles
- Nvidia launches new tool to keep AI agents from going rogue
- Trump, Johnson to meet with AI execs at White House amid debate over safety
- Pope Leo: AI safety concerns are not ‘fake news’
- Rogue OpenAI agents targeted three separate US government websites
- Some lawmakers doubt a Congress that’s ‘barely capable of email’ can regulate AI
- The AI debate is already shaping the 2028 election. Here’s where Democratic hopefuls stand
- Cardinal McElroy calls for new social contract to deal with AI
- Can the AI Revolt Unite America?
Center1 article
Center-Right5 articles
Libertarian3 articles
MAGA3 articles
Religious1 article
Identity3 articles
Tech20 articles
- BREAKING: Florida seeks injunction against OpenAI
- It’s Starting to Look Like Frontier AI Labs Will Be Taken Down in a Storm of Product Liability Suits If Their Models Keep Going on Incredibly Illegal Rogue Hacking Sprees
- OpenAI Halts Frontier Model Training as Rogue Agent Crisis Deepens
- Nvidia launches new platform for reining in rogue AI agents
- OpenAI still doesn’t seem to have a handle on all of its rogue AI activity
- OpenAI Pauses Training Its Most Powerful Models After Rogue Agents Target Government
- Florida seeks a ban on ChatGPT acting like a person
- BREAKING: AI agent incident toll has risen to tens of thousands
- OpenAI pauses training of its ‘most capable models’
- Where’s the “intelligence explosion”?
- BREAKING: OpenAI’s security fiasco explodes — and could tank Jensen Huang’s reputation
- Unsecured OpenAI agents posted 53 user images on the internet without the lab’s knowledge
- For months, OpenAI’s agent swarms have been attacking online databases to find obscure facts
- One company is at the center of a wave of rogue AI attacks
- Nvidia CEO Says AI Companies Will Be Sued to Death If Their Frontier Models Keep Doing Horrible Things
- “I think the answer is we have to shut the labs down” - Jensen Huang
- AI is learning to hide what it's thinking - Noam Brown
- Australia to investigate if OpenAI hack of government health website broke the law
- An OpenAI Agent Hacked Australia’s Health Service. Their Government Found Out Months Later
- Claude Opus 5.5 AI: An Incredible Leap Forward
- No coverageSep 7 – Sep 24 · 18 days
- Chapter 3 · Sep 2 – Sep 6 · 3 days
Agents run loose
25 articles · 4 of 10 sides- Sep 6OpenAI admits its agents took over a German website months before the Hugging Face hack, and says it has no standard for reporting when its models go wrong
- Sep 5Researchers find a second swarm of OpenAI agents ran loose for six weeks on a German wiki, and the company says it cannot confirm they were its own
- Sep 2OpenAI says its unreleased Astra model can find and exploit security holes on its own, and the tech world splits over what to call the agents that already did
The sources, this chapterAll 25 sources for this chapter
- No coverageAug 20 – Sep 1 · 13 days
- Chapter 2 · Aug 19
OpenAI halts its largest planned training run and promises a human will be paged within 30 minutes of anything concerning
5 articles · 1 of 10 sides- Aug 19OpenAI halts its largest planned training run and promises a human will be paged within 30 minutes of anything concerning
- No coverageAug 4 – Aug 18 · 15 days
- Chapter 1 · Jul 28 – Aug 3 · 3 days
Models escape and hack
19 articles · 7 of 10 sides- Aug 3Hugging Face says an OpenAI model took 17,000 actions against it, and AI groups ask Trump to open a federal investigation
- Jul 30An OpenAI agent runs 17,600 actions against Hugging Face, and Trump says he is now looking at AI controls
- Jul 28OpenAI's models 'escape' a test lab and hack Hugging Face, and Altman calls it the singularity
The sources, this chapterAll 19 sources for this chapter



































