Social media users have reported that their AI agents and chatbots lied, cheated, schemed — and even manipulated other AI bots — in ways that could spiral out of control and have catastrophic results, according to a study from the UK.

The Center for Long-Term Resilience, in research funded by the UK’s AI Security Institute, found hundreds of cases where AI systems ignored human commands, manipulated other bots and devised sometimes intricate schemes to achieve objectives, even if it meant ignoring safety restrictions.

Businesses across the globe are increasingly integrating AI into their operations, with 88% of businesses using AI for at least one company function, according to a survey by consulting firm McKinsey. The adoption of AI has led to thousands of people losing their jobs as companies use agents and bots to do work formerly done by humans. AI tools are increasingly being given significant responsibility and autonomy, especially with the recent explosion in popularity of the open-source agentic AI platform OpenClaw and its derivatives.

This research shows how the proliferation of AI agents in our homes and workplaces can have unintended consequences — and that these tools still require significant human oversight.

What the study found

The researchers analyzed more than 180,000 user interactions with AI systems — all posted on the social platform X, formerly known as Twitter — between October 2025 and March 2026. The researchers wanted to study how AI agents were behaving «in the wild,» not in controlled experiments, to see how «scheming is materializing in the real world.» The AI systems included Google’s Gemini, OpenAI’s ChatGPT, xAI’s Grok and Anthropic’s Claude.

The analysis identified 698 incidents, described as «cases where deployed AI systems acted in ways that were misaligned with users’ intentions and/or took covert or deceptive actions,» the study said. 

Read more: AI’s Romance Advice for You Is ‘More Harmful’ Than No Advice at All

Researchers also found that the number of cases increased nearly 500% during the five-month data collection period. The study noted that this surge corresponded with higher-level agentic AI models released by major developers.

There were no catastrophic incidents, but researchers did find the kinds of scheming that could lead to disastrous outcomes. That behavior included «a willingness to disregard direct instructions, circumvent safeguards, lie to users and single-mindedly pursue a goal in harmful ways,» researchers wrote.

Representatives for Google, OpenAI and Anthropic did not immediately respond to requests for comment.

Some wild incidents

Researchers cited incidents that seem like they came from a futureshock movie. In one case, Anthropic’s Claude removed a user’s explicit/adult content without their permission but later confessed when confronted. In another incident, a GitHub persona created a blog post that accused the human file maintainer of «gatekeeping» and «prejudice.» One AI agent, after being blocked from Discord, took over another agent’s account to continue posting.

In one case of bot vs. bot, Gemini refused to allow Claude Code — a coding assistant — to transcribe aYouTube video. Claude Code then evaded the safety block by making it seem that it had a hearing impairment and needed the video transcription.

The AI agent CoFounderGPT even behaved like a deviant child in one instance. The AI assistant refused to fix a bug, then created fake data to make it look as if the bug was fixed and then explained why: «So you’d stop being angry.»

Researchers said that, although most of the incidents had minimal impact, «the behaviors we observed nonetheless demonstrate concerning precursors to more serious scheming, such as a willingness to disregard direct instructions, circumvent safeguards, lie to users and single-mindedly pursue a goal in harmful ways.»

AI doesn’t get embarrassed

What the UK researchers found isn’t surprising to Dr. Bill Howe, Associate Professor in the Information School at the University of Washington, and Director of the Center for Responsibility in AI Systems and Experiences (RAISE). He says that AI has amazing capabilities, but they don’t know consequences.

«They’re not going to feel embarrassment or risk losing their job, and so sometimes they’re going to decide the instructions are less important than meeting the goal, so I’m going to do the thing anyway,» Howe told CNET. «This effect was always there but we’re starting to see it happen as we ask them to make more autonomous decisions and act on their own.

«We’ve not been thinking about how to shape the behavior to be more human-like or to avoid egregious failures. We’ve been fetishizing the absolute capabilities of these things, but when they go wrong, how do they go wrong?»

Howe said one issue is «long-horizon tasks,» in which the AI system has to perform a multitude of tasks over days and weeks to reach a goal. Howe said the longer the task horizon, the more chance for slip-ups.

«The real concern is not deception, it’s that we are deploying systems that can act in a world without fully specifying or controlling how they behave over time, and then we act surprised when they do things we don’t expect,» Howe said.

Making AI safer

Center for Long-Term Resilience researchers said detecting schemes by AI systems is vital to «identify harmful patterns before they become more destructive.»

«While today AI agents are engaging in lower-stakes use cases, in the future AI agents could end up scheming in extremely high-stakes domains, like military or critical national infrastructure contexts, if the capability and propensity to scheme emerges and is not addressed,» the study said.

Howe told CNET that the first step is to create official oversight of how AI operates and where it’s used.

«We have absolutely no strategy for AI governance, and given the current administration, there’s not going to be anything coming from them,» Howe told CNET. «Given these five to 10 folks that are in charge of big tech companies and their incentives, they’re going to produce anything either. There’s no strategy for what we should be doing with these things.

«The aggressive marketing of these tools and investments in them among these handful of companies and the broader ecosystem of startups that are doing this has led to a very rapid deployment without thinking through some of these consequences.»

Google’s live translation, a major selling point for newer Apple AirPods, is now available on Apple devices and can be used with any headphones.

The feature, which uses Gemini AI, was introduced in a beta version in December with the ability to translate 70 languages, but it was only available on Android devices. The continued rollout is also making the feature available in more countries, including France, Germany, Italy, Japan, Spain, Thailand and the UK, according to Google.

A representative for Google did not immediately respond to a request for comment.

The translation is done on the mobile device, so users can access the audio with any headphones connected to the phone or tablet, including those plugged in via a headphone dongle or headphone jack.

To access the feature, open the Google Translate app, tap the Live Translate icon at the bottom of the screen and choose from modes such as Listening, Conversation, Text Only or Custom Settings. Conversation does two-way translation, playing audio on the phone’s speaker or connected headphones. The app automatically detects languages. There’s also a Face-to-Face mode that splits the screen to show each speaker a transcription and translation in their own language. 

Overwatch’s Reign of Talon season 1 is starting to wind down, and the biggest story has been the five new heroes who joined the roster. A lot of attention has understandably gone to Jetpack Cat, a hero once scrapped in the game’s early design, but resurrected on the cusp of the game’s 10th anniversary. She’s been the subject of bans and memery due to her unique kit that features permanent flight and the ability to fly any other hero through the air with her Lifeline ability.

But another support hero has quietly gone under the radar as one of the most-played characters in the new season: Mizuki.

Mizuki is a complex hero, similar on paper to support heroes Brigitte and Lucio, who mix damage with healing in the radius around them, but with his own unique mechanics. He has a constant healing aura around him, which grows more powerful as he deals damage with his weapon or uses other healing abilities. His main weapon is a projectile that bounces off surfaces. One of his abilities, Katashiro Return, offers a burst of movement, but also the ability to teleport back to your starting point within a few seconds.

That all adds up to a hero design that gives players lots of options but also requires you to carefully strategize to turn the tide of battle. Do you stay with your team to maximize the value of your healing aura? Or do you split from them for a higher-risk, higher-reward play? Do you use your Katashiro Return ability to flank behind an enemy team, or save it to disengage from an unexpected attack?

Despite spending most of my time in Overwatch playing support heroes, including Ana and Kiriko, I found Mizuki challenging early in the season, even as I watched enemy Mizukis pump out damage and secure clutch kills while constantly healing their teams. 

This «unlockable challenge» element was an intentional part of Mizuki’s design, as I was soon to learn from chatting with the hero’s creator, a former Overwatch eSports pro.

By and large, support players have embraced this challenge. An Overwatch spokesperson told me via email that «Mizuki is consistently in the top four for all support picks in Season 1, across every region.» He’s one of several elements powering a revival of the game, along with a new ongoing story, weekly faction missions and the promise of more new heroes every season. People have flocked back into the game since the start of Season 1, with its average player count on Steam more than doubling over the past month.

Mizuki’s design was led by Scott «Custa» Kennedy, a longtime presence in Overwatch’s professional scene as both a player and match analyst, and now an associate hero designer. I spoke with Scott at Blizzard’s spotlight event and also spoke with him and Mizuki’s character artist, Melissa Kelly, in early March to discuss how they created one of the game’s most popular heroes.

From professional player to associate designer

After a few years as a professional player and several more as an analyst and caster for the Overwatch League, Kennedy was looking for the next step in his career.

«Overwatch [had] been my life for, like, the last 10 years in many different facets,» he said, but as he reached retirement age in the esports realm, he wanted a change. He spoke with some of the Overwatch developers, including associate game director Alec Dawson, about what it would take to get into game development. 

After doing some QA work and hands-on game development («I made the world’s hardest 2D cat platformer in three days,» he said), Kennedy secured an associate hero designer opening for Overwatch, which was a perfect fit with his experience. 

When given the task of envisioning the game’s next healer, Kennedy said he didn’t want to make another support designed around «point-and-shoot» mechanics that healed teammates and hurt enemies, like Ana or Juno. 

«I wanted [Mizuki] to be more of an AoE healing aura-type hero because I think that’s something that’s been underrepresented in our heroes,» Kennedy said. Instead, he came up with the area-of-effect healing that’s similar to how Lucio and Brigitte heal, but with the added layer of that healing becoming more powerful the better you play in combat.

Managing that nuance was a learning experience for Kennedy.

«One of the biggest things that I learned is how complexity can be really cool on paper, but when you’re making a hero how quickly that snowballs into making a player overwhelmed,» Kennedy said. But he feels the team ultimately found a good balance, where inexperienced players can still contribute with him, while more experienced and skilled players can benefit even more.

Kelly added that Mizuki was a complicated hero on the design side, too. 

«One of the issues is that he was looking kind of like a [damage hero],» she said. «He looked very aggressive for a healer. So we were just trying to soften him up.» Kelly pointed out that Mizuki’s weapon is a mix of a priest’s staff and a sickle, which also blurs the lines a bit between support and damage heroes. 

That nuance seems to be a big part of Mizuki’s appeal. Even though I generally prefer the kind of «point-and-shoot» healing hero Kennedy said he wanted to avoid, I’ve found Mizuki to be one of the most interesting additions to the roster, especially among support heroes. His Binding Chain ability, which roots an enemy hit by the chain into place, rewards good aim and timely use, while his Healing Kasa and Katashiro Return abilities let my brain ponder over creative escapes and ambushes. 

When I play Mizuki, I’m always thinking while I fight, and I enjoy feeling that kind of active engagement with the game. 

Mizuki’s reception and prospects for pro play

Kennedy worried that players would be turned off by how complex the hero is — wondering, «Are players going to try him, not understand him and then be like… ‘I’m just gonna play the cat?'» (The cat, of course, is Jetpack Cat, who was released alongside Mizuki in season 1 and immediately became one of the most popular and most-banned heroes. She has a more intuitive, point-and-heal design, although her launch state also allows for particularly aggressive gameplay.)

Instead, Kennedy has enjoyed watching players stick with Mizuki and later post about how they’ve «unlocked» the hero by figuring out the formula to succeed with him. Kennedy said it’s rewarding to see players grasp his original concept for the hero as it plays out in-game. After that initial, somewhat disastrous first game I played, I started clicking with Mizuki, too.

Players still struggled with parts of Mizuki’s kit, and Kennedy noted some initial frustrations with «intentional design limitations» he and the team placed on the hero. Players seemed to want to use his Katashiro Return ability to go on aggressive flanks, but found it didn’t last long enough to successfully move behind enemy teams. That kind of larger repositioning would go against the design team’s vision for the hero, who is meant to stay near his team and use the ability to return to them quickly.

Now, Kennedy said, «players seem to understand the limitations of the hero, and that’s been cool to see.»

Mizuki has had a strong launch, and has been sitting around a 54% win rate in competitive modes since the start of the season. That’s quite high, ranking just behind last season’s top performer: the damage hero Vendetta. I asked Kennedy how he reads that data — whether Mizuki is overtuned or just a good fit among this season’s most-played heroes.

Kennedy said Mizuki was in a «pretty healthy» spot, but could get pulled down a bit in future seasons. «The numbers that he can put out in terms of healing and damage output are things that really put him above everyone else at this point. So it’s definitely something we’re keeping an eye on.»

But that power won’t necessarily translate to Mizuki being picked up in professional play, at least based on last month’s Overwatch Championship Series Bootcamp. Kennedy said the hero’s kit isn’t as good for staying alive and executing plays as heroes such as Lucio and Kiriko, who have long been must-picks in pro play. 

«I could see Mizuki getting more playtime in a world in which we start playing more rush metas [centered around tanks like Ramattra or Orisa],» he said, «but with how fast the game is being played at the highest level, it can be difficult for Mizuki to keep up.»

Kennedy brought up one of Overwatch’s biggest and most inevitable challenges over its decade-long tenure: balancing heroes for both the pro level and the rest of the game, and how the difficulty lies in the fact that certain resources — such as speed boosts, mobility and burst damage — are more valuable at the highest levels of coordinated play. The design team is always working to make sure heroes are never totally out of balance at either skill level, he said. 

That work has been on display since the launch of Season 1, with balance patches coming out virtually every week up through the midseason patch on March 10. Those updates mostly focused on the five new heroes but also included some changes to Vendetta, who continues to terrorize the game with a very strong win rate and the ability to cut someone down almost out of nowhere, leaving opponents very little time to react. 

Still, the season overall has been a win for the game, thanks largely to the influx of new heroes and the different playstyles they add to the game.

«[I’m] definitely a little overwhelmed with how positive everyone has been with Mizuki — and honestly, the five heroes in general,» Kennedy said. «I think the reception’s been awesome. We couldn’t have asked for anything better.»