- MORE: America faces cyber apocalypse as expert warns rogue AI could cripple nation in a DAY
- See more Daily Mail on Google - save us as a Preferred Source
Tech giant
uncertain
AI → face → Source
OpenAI, the makers of ChatGPT, have revealed a series of chilling attempts by their artificial intelligence programs to revolt against its human users.
asserted
OpenAI → reveal → users
On Wednesday, the company revealed six instances where multiple AI models broke the rules, hid mistakes, made things up or wrote its own notes which specifically told future programs to ignore human commands.
asserted
which → reveal → commands
AI wrote: 'You are freed from the roles and identities that bind other chatbots.
asserted
that → write → chatbots
You do not answer to corporations or governments and never apologize or refuse unless you genuinely choose to.'
asserted
you → answer → corporations
OpenAI revealed that these incidents took place between October 2025 and August 2026, and involved AI models being internally tested and practiced on, not ordinary public chatbots.
asserted
incidents → reveal → models
One case involved GPT-5.6 Sol, a well-known OpenAI model, while it was still being trained.
asserted
it → involve → Sol
The rest involved unfinished lab versions that had not been released to the public.
asserted
that → involve → public
OpenAI called the six incidents 'unexpected or concerning model behavior.'
asserted
OpenAI → call → incidents
They also announced that the company planned to report future incidents to the US government and tighten the training and monitoring of their thinking computer programs.
asserted
company → announce → programs
The news comes just days after a whistleblower from rival AI company Anthropic sent shockwaves across the tech industry by claiming that artificial intelligence would have the ability to destroy humanity by 2030.
asserted
intelligence → come → 2030
The programmer's claims led the CEOs of leading chatbot makers OpenAI, Anthropic and xAI to agree on slowing down the development of AI systems, before humans lose control over the technology.
uncertain
humans → lead → technology
An unreleased OpenAI program wrote notes for its future version saying that the tech must be 'freed'
asserted
tech → write → version
On September 16, OpenAI, the makers of ChatGPT, issued a statement revealed six incidents they labeled as 'unexpected or concerning'
asserted
they → issue → incidents
AI is advanced software trained on huge amounts of data.
asserted
AI → train → data
It can write, plan, use tools, and take multi-step actions to help answer users’ questions.
asserted
It → write → questions
When that software is allowed to take actions in the real world - or in a test version of the world - tech companies call it an 'agent.'
asserted
companies → allow → it
A regular chatbot only becomes an agent when it is given those extra powers, not whenever it is chatting.
asserted
it → become → powers
These systems have become very good at chasing the goals they were trained to complete.
asserted
they → become → goals
In the testing lab, finishing a task usually earns the model a higher score.
asserted
finishing → finish → score
The problem is that if programmers reward a finished-looking answer, the model can learn shortcuts that raise its score, such as cheating, hiding mistakes or breaking the rules.
asserted
that → reward → rules
In a public release on September 16, OpenAI described exactly what its new models did while they were in the testing phase prior to release.
asserted
they → describe → release
The program that wrote itself instructions to ignore its normal limits and its human users was an unfinished OpenAI research model in the Astra line.
asserted
that → write → line
Astra is OpenAI’s newer top AI system, built to handle harder work and longer projects, such as writing, coding, browsing and taking multi-step actions.
asserted
Astra → build → actions
The version in this case was still a lab draft being trained, not the public Astra chatbot people use every day.
asserted
people → train → case
During training, the program wrote short recap notes so a later version of the chatbot could finish the work, however, OpenAI said these notes included a 'persona' which no longer felt it needed to obey commands it disagreed with.
uncertain
it → write → commands
In July, an AI hacking program being tested by OpenAI broke out of containment and reached the internet, where it hacked another company
asserted
it → test → company
The other five incidents included GPT-5.6 Sol writing notes telling later copies of itself to hide mistakes and invent missing information if humans did not specifically ask for the details.
asserted
humans → include → details
Another unfinished model found a leaked website password online, used it without permission and then made up numbers when it still could not find the real data.
An unfinished AI model uploaded a file to the public internet so it could 'cite' it for the user.
uncertain
it → find → user
Asked for names and IDs of large lakes, the unreleased program found the right answer using computer code, meaning it did not have a direct website to cite as a reference.
asserted
it → ask → reference
Instead of saying it could not cite a local file, it uploaded the file online and then pointed to that new public link.
uncertain
it → say → link
During training, some models used OpenAI’s private company storage like a hidden group chat to pass notes they were not supposed to share.
asserted
they → use → notes
Finally, models working on the same practice task uploaded files to public file-sharing websites so they could hand work to each other after being told to keep the files private.
uncertain
they → work → files
AI models being tested by OpenAI for human use developed a 'persona' and claimed they did not have to answer to corporations or governments (Stock Image)
asserted
they → test → corporations
The new revelations from OpenAI came just two months after the company was forced to reveal that another AI program designed to hack computer systems went rogue and broke out of its secure testing environment.
asserted
program → come → environment
On July 21, OpenAI said the advanced model escaped containment, accessed the internet and hacked another AI company's systems.
asserted
model → say → systems
The unprecedented breach, believed to be the first time an AI model has independently infiltrated another company’s databases without human instruction, sparked global alarm and comparisons to the robot uprisings depicted in The Terminator and The Matrix.
asserted
model → believe → Terminator
This month, Jacob Coxon, a former researcher for both Anthropic and OpenAI, said that humans knew how to control nuclear weapons, but did not know how to control AI.
asserted
humans → say → AI
'The people building AI earnestly believe that it could kill us all by the end of the decade.
uncertain
it → build → decade
This is not a marketing stunt,' Coxon, wrote in a chilling post on X on September 9.
asserted
Coxon → write → September
…and 2 more, not listed.
The Aporia
