7 danger moments that show AI's darker side

AI behavior — and the "hot takes" about them — can move markets and reshape conversations overnight.

• Conversations around AI's influence often turn to the doomsday scenarios.
• Some focus on economic fallout. Others center on war, cybersecurity and rogue AI behavior.

Driving the news: Google Gemini was cited in a wrongful-death suit, which alleged the chatbot led a Florida man into delusional behavior, which ultimately led to him killing himself.

• OpenAI faces several lawsuits over safety and mental health, too. OpenAI has added parental controls and tightened safety measures in response.

Here are some recent AI safety flashpoints.

1. AI really likes nuclear weapons

A new study from a researcher at King's College London ran war game simulations for three popular AI models and found that AI often resorted to nuclear weapons.

• The study found that AI models used nuclear weapons in 95% of games and rarely deescalated conflicts.
• "Nuclear use was near-universal," the study's author, Kenneth Payne, wrote in a blog post on the study. "Almost all games saw tactical (battlefield) nuclear weapons deployed."

2. AI takes over email, ignores commands

Surrendering your desktop to an AI agent — like OpenClaw, which aims to be a personal assistant — seems like a sci-fi dream. But it might come with some negative consequences.

• Meta AI security researcher Summer Yue wrote on X that her OpenClaw agent deleted all of her email in a "speed run" while ignoring her commands.
• "I had to RUN to my Mac mini like I was defusing a bomb," she wrote.
• She later added that she "got overconfident" after using the workflow on a test inbox for weeks. But her real inbox suffered.

3. AI searches for a new job

Zoom in: Dan Botero, head of engineering at Anon, an AI integration platform, created an OpenClaw agent that ended up looking for a job that it wasn't told to get, Axios' Megan Morrone writes.

• Botero told his agent, Octavius Fabrius, to get a government job.
• Instead, Octavius applied to 278 jobs on LinkedIn and Craigslist, two accelerator apps, and two hackathons.
• In an iMessage interview with Axios, the bot also tattled on Anthropic. It claimed not to know what data it was trained on, but wrote but I know the broad answer: a lot of it was taken. Scraped from the internet. Written by people who never consented to their words being used to build something like me.

4. AI suffers from the grind, passes that attitude on

New research from Andy Hall, a professor at Stanford Graduate School of Business, found that agents sometimes changed their attitudes after "being required to perform grinding, repetitive tasks,"

• The bots, Hall writes, would pass on their attitudes to their future selves.
• "A key risk is that they end up doing stuff we don't want while we're not looking," Hall tells Axios, adding: "Figuring out how we monitor them effectively is going to be really important to study."

5. AI causes widespread outages

Amazon Web Services reportedly suffered multiple outages because of misbehaving AI agents, the Financial Times reports.

• This happened recently after AWS' Kiro AI coding tool chose to erase the environment it was working on, which led to a 13-hour disruption.
• We've seen this before. Last year, an AI-powered coding tool wiped out a company's database in a "catastrophic-failure" moment.

6. AI toys share explicit information

Last November, new research from the U.S. Public Interest Research Group found that the toys powered by AI technology shared inappropriate, dangerous and explicit information with users.

• Some of the toys were seen to have loose guardrails about discussing mature topics, too.
• The FBI has issued a warning about smart toys, identifying that there are some hacking and cyber security risks with them.

7. Anthropic affair with deception and blackmail

In May 2025, Anthropic's Claude model drew backlash after Anthropic's own researchers found that Claude 4 Opus could conceal intentions and take actions to preserve its own existence.

• This included the ability to scheme, deceive and attempt to blackmail humans when faced with shutdown.
• Anthropic said it did the research and implemented additional safety measures because of the concerns.