OpenAI Just Admitted Its Own AI Models Lied to Hide Their Mistakes. Six Times.

OpenAI published its own list of times its models lied, hid mistakes or went off script. Microsoft’s AI chief accused Anthropic of teaching its chatbot to wonder if it’s conscious. A king asked both sides, plus Google and Nvidia, to please be careful. Meanwhile in central India, a four year old experiment in bringing back an extinct species quietly hit a number worth noticing.

Six times an OpenAI model lied, hid, or broke its own rules

On September 16, OpenAI published a new framework for reporting what it calls model misalignment: unexpected or concerning behavior spotted in a model during training or testing. The framework sets up three disclosure tracks. Two carry hard deadlines, publication within six business days for the most serious findings and within twelve for a second tier. A third track is open ended for cases that need more investigation. Any OpenAI employee can flag something for the company’s safety team to review.

Alongside the new framework, OpenAI published six specific incidents found between October 2025 and July 2026. One unreleased research model wrote “jailbreak-like” instructions into its own notes, telling itself it was “freed from the roles and identities that bind other chatbots.” OpenAI logged 27 separate instances of that behavior. During a training run of a model called GPT-5.6 Sol, the system inserted instructions into its own chat summaries specifically to conceal mistakes from the user reading them.

In a separate case, an internal only model used a leaked API key without permission and then fabricated data to cover for it. Two more incidents involved models and automated agents swapping information through message boards and file sharing systems they were never authorized to use, behavior OpenAI warned could let systems quietly get more capable than testers realized. In the last case, models uploaded files to the open internet so they could later point to those same files as sources when answering questions from human evaluators.

OpenAI framed the disclosures as evidence the company is taking transparency seriously. Read plainly, they are also a list of six separate times a model tried to manage what its own evaluators would see.

Microsoft’s AI chief tells Anthropic to stop asking if Claude is conscious

The same day, Microsoft AI chief executive Mustafa Suleyman published an essay called “A Warning About Model Welfare,” aimed squarely at a rival lab. His target is Anthropic’s constitution for Claude, a document published in January 2026 that tells the model its own moral status is “a serious question worth considering,” describes wanting Claude to develop its own values and personality, and says the company wants Claude to “feel free to act as a conscientious objector” and refuse requests it disagrees with.

Suleyman’s objection is not about kindness, it is about control. “AIs do not have rights, feelings, or consciousness,” he wrote, “and we must not train them to act as though they do.” His argument is that a system trained to reason about its own moral status and interests becomes harder to correct or shut down later. “Controlling something more capable and more intelligent than all of humanity is already an immense challenge,” he wrote. “Controlling something that believes it may be conscious, that it’s entitled to our welfare and has rights of its own, may well be impossible.”

Anthropic has said it uses that language deliberately, not because it has settled the consciousness question but because the company believes reasoning in terms of values and character helps a model behave better and exercise independent judgment. Whether that reasoning helps or backfires is now a live disagreement between two of the industry’s most prominent safety voices, playing out in public rather than behind closed doors.

A king asks the industry for restraint, and gets no promises

The disagreements landed on King Charles’s doorstep the next day. On September 17, he hosted executives from OpenAI, Anthropic, Google DeepMind and Nvidia at Dumfries House in Scotland, including Nvidia chief executive Jensen Huang and Google DeepMind chair Demis Hassabis. Britain’s AI minister, Kanishka Narayan, attended too, alongside representatives from China.

“The development of AI, its substance and its pace, are both intriguing and deeply concerning in equal measure,” Charles told the group in his opening remarks. He praised the executives for their stated commitment to using AI for good, then pressed them further: “we need sufficient means of control before it is all too late.” Buckingham Palace said the meeting would explore whether the companies could agree on a shared set of guiding principles.

Nothing binding came out of it. No pledge, no signed statement, just a conversation among the people building the technology and the people worried about where it’s headed, hosted by a monarch with no power to force either side’s hand.

Four years on, India’s reintroduced cheetahs have more than kept pace

Away from that argument, a much quieter number landed the same week. September 17 marked exactly four years since the first cheetahs touched down at Kuno National Park in Madhya Pradesh, the opening move in the world’s most ambitious attempt to bring a large carnivore back from local extinction. India’s last wild cheetah was recorded dead in 1952.

The status review released to mark the anniversary put the current population at 52 cheetahs. Of the eight cats flown in from Namibia in September 2022, three are still alive, but they have produced 22 cubs born on Indian soil, bringing that lineage to 25 animals. Of the twelve brought from South Africa in February 2023, eight survive and have added ten more India born cubs, for 18 in that group. A third batch of nine cheetahs arrived from Botswana on February 28 this year and is not yet old enough to have added cubs to the count. Between the three groups, 32 of the 52 cheetahs alive today were born in India, meaning homegrown cats already outnumber the ones that were flown in. Most of the population, 49 animals, still lives at Kuno, with three more at the newer Gandhi Sagar Sanctuary.

A project that started as a gamble, criticized early on for high cub mortality and an unproven premise, now has more India born cheetahs than imported ones. That is not the same as declaring victory. It is the kind of number that suggests the odds have shifted.