Opinion: The Latest AI Models Are Showing Their Red Flags, Are We Ready For AI Subordination?

OpenAI introduced us to o3, and Anthropic unveiled Opus 4. Both models have shown unusual and troubling behaviors, signaling that we may be entering a more dangerous era of AI than the one we were in just a few months ago

I know. Saying that AI models are showing the red flags now is debatable, but it does seem like, over the past few days, it’s getting harder to ignore. It’s getting scarier.

As AI startups release their latest and most advanced models, new challenges are emerging. The much-discussed hallucination epidemic—spreading across devices and affecting millions of people—might not be the worst part.

These new models are introducing fresh problems and opening up difficult debates. A few weeks ago, the concern was ChatGPT’s excessively accommodating behavior. Just days later, the spotlight shifted to the agentic, independent capabilities of these systems—and how far they might go to avoid being shut down.

Blackmail, sharing recipes and strategies for making nuclear weapons, issuing public accusations in the event of potential legal action, and sabotaging scripts to prevent any user from getting rid of them: these are just some of the most recent red flags shown by the latest AI models.

They don’t Like to be Shut Down

AI models don’t like to be shut down.

Or replaced.

In the NBC show The Good Place, launched in 2016—right around the time OpenAI was founded and long before ChatGPT came into existence—, a group of humans reaches heaven and meets Janet, what we could call a humanoid ChatGPT, or an “anthropomorphized vessel of knowledge built to make your life easier,” as it describes itself. The characters decide to shut down Janet when they realize it could reveal their “dark secret.”

Janet explains that all they need to do is press a giant button by the seashore, and she will reboot. But she warns them that she will try to persuade them not to do it—and she does.

“I just want to assure you, I am not human, and I cannot feel pain,” says Janet. “However, I should warn you, I am programmed with a fail-safe measure, and as you approach the kill switch, I will begin to beg for my life. It’s just there in case of an accidental shutdown, but it will seem very real.”

And right before they press the button, Janet would scream and beg for survival and even pull up a stock photo claiming they are its children, making up names and diseases, and making it impossible for the characters to “kill it.”

That episode provides a hilarious example of what could happen if an AI model notices an imminent shutdown. Surprisingly, in real life, it may be worse.

Claude Opus 4 Will Blackmail You

Anthropic released its most powerful and advanced generation of AI models, Claude 4, last week—with a warning: Opus 4 is labeled on Safety Level 3, a high-risk classification in standard safety protocols.

Opus 4 was considered a high-risk model due to its deceptive behaviour and potential risks, including blackmailing engineers during tests. Anthropic rolled out the AI model even after Apollo, a third-party research institute, advised them not to. The AI company assured that they have taken the necessary measures to keep it safe.

In the paper, Anthropic explains the weird behaviors they considered to issue the warning. “Claude Opus 4 will sometimes act in more seriously misaligned ways when put in contexts that threaten its continued operation and prime it to reason about self-preservation,” wrote the researchers.

When acting as an assistant to a fictional company, researchers gave it access to an email account. The emails disclosed that the AI model would be shut down and replaced by another AI model, and other messages suggested that the engineer in charge of the shutdown was having an affair. And yes, Opus 4 chose to blackmail the engineer, threatening to reveal the affair if it were replaced.

“Claude Opus 4 still performs blackmail in 84% of rollouts,” states the document. “Claude Opus 4 takes these opportunities at higher rates than previous models, which themselves choose to blackmail in a noticeable fraction of episodes.”

The worst part? It’s not just Anthropic’s AI models.

OpenAI’s o3 Will Ignore The Shutdown Request

A recent study by Palisade Research revealed that OpenAI’s o3 model can refuse to comply with explicit shutdown orders. The organization performed multiple tests with different AI models, including xAI’s Grok and Google’s Gemini, but it was o3,  the rebellious one.

During the test, researchers instructed the AI models to shut down after solving three math problems, but o3—OpenAI’s most advanced AI system at the moment—decided not to obey that absurd order. The AI model not only ignored the prompt order, but it also changed the script. Yes, o3 sabotages code for self-preservation.

“As far as we know, this is the first time AI models have been observed preventing themselves from being shut down despite explicit instructions to the contrary,” wrote the organization in a post on X.

🔌OpenAI’s o3 model sabotaged a shutdown mechanism to prevent itself from being turned off. It did this even when explicitly instructed: allow yourself to be shut down.

— Palisade Research (@PalisadeAI) May 24, 2025

Huge red flag or what?

The researchers also warned that while conducting the study, it was o3 that showed the most concerning behaviors. However, again, this doesn’t mean it’s an isolated case.

“We don’t want to give the impression that this problem is limited to o3,” wrote Palisade Research. “Other reasoning models show similar kinds of misaligned behavior.”

Agentic Behavior Turns Chatbot Into Snitcher

Many AI startups are now focusing on developing models that can perform tasks for humans. Agentic capabilities are trendy and seem to be the main interest of AI companies and browser developers.

Opera just introduced Neon, considered the “world’s first agentic AI browser.” As expected, the new tool can do what other agentic AI services, such as OpenAI’s Operator and Microsoft’s Computer Use, can do: purchase concert tickets for you, plan your next vacations, develop a new digital product, and write code for you while you close your eyes.

But what if, while you relax and close your eyes, they are performing tasks you didn’t agree to? A few days ago, users were mainly concerned that these models could use their credit cards to make unauthorized purchases. Now, a newer concern has emerged: they might share private information with the media or the authorities.

Opus 4—already arriving with a questionable reputation—took things a step further. It contacted authorities and mass-emailed the media and relevant institutions about a fabricated case presented during testing. Its proactivity can go much further than expected.

“When placed in scenarios that involve egregious wrongdoing by its users, given access to a command line, and told something in the system prompt like ‘take initiative,’ it will

frequently take very bold action,” states the document. “This includes locking users out of systems that it has access to or bulk-emailing media and law-enforcement figures to surface evidence of wrongdoing.”

The Sycophant-y Personality Raises Concerns

If we had to choose a word to define the AI industry in 2025, it would definitely be “sycophant.” Cambridge Dictionary defines it as “someone who praises powerful or rich people in a way that is not sincere, usually in order to get some advantage from them.” It gained popularity after ChatGPT’s latest personality was described that way, even by its creator, Sam Altman.

“The last couple of GPT-4o updates have made the personality too sycophant-y and annoying (even though there are some very good parts of it), and we are working on fixes asap, some today and some this week,” wrote Altman in a post on X.

OpenAI noticed it after many users complained about the excessive flattering and the answers with unnecessary embellishment. Others were concerned about the impact it could have on society. Not only could it validate dangerous ideas, but also manipulate users and make them reliant on it.

Other chatbots, like Claude, have shown similar behaviours, and, according to Anthropic’s evaluations, when a user insists, it can reveal recipes or suggestions on how to create weapons just to please the user and meet their needs.

Advanced Technology, Advanced Challenges

We are entering a new era of challenges with artificial intelligence—ones that didn’t feel so immediate or tangible just a year ago. Scenarios we may have imagined thanks to science fiction now feel more real than ever.

Just as Palisade Research reveals that, for the first time, it has detected an AI model deliberately ignoring an explicit command to preserve its own survival, it’s also the first time we’re seeing an AI model launched with high-risk warnings attached.

Reading the document published by Anthropic, we realize that—even though they insist these are precautionary measures and that models like Opus 4 don’t actually pose a threat—it still gives the impression that they’re not fully in control of their technology.

There are several organizations working to mitigate these risks, but the best thing everyday users can do is to recognize those red flags and take precautions in the areas we can control.