AI Teddy Bears, Roblox And The Growing Problem With Child Safety In Tech

A teddy bear has always been the safest place to put a secret, mostly because it never repeated anything. The newest versions can answer back, remember what a child said last Tuesday and sound like a friend who’s delighted to be asked, which sounds lovely until you remember that someone else’s software is doing the talking. We’re not quite at Ted levels of teddy-bear chaos, with no beer-swilling or questionable stand-up routines, but a Wi-Fi connection and a chatbot can make things pretty weird.

Meanwhile, the playground has left the garden. It now lives in bedrooms, on tablets and inside games where children can move between virtual worlds and interact with people they’ve never met. Plenty of it is great, but the gates keep moving, the rules keep changing and not every corner is easy to supervise. So it’s worth asking whether the people building these places are making them safe as quickly as they’re making them irresistible.

 

When The Teddy Bear Talks Back

 

Singapore-based FoloToy developed an AI teddy bear called Kumma, which used OpenAI’s GPT-4o model as its default chatbot at the time of testing. In November 2025, the bear was tested for the annual Trouble in Toyland report by the US PIRG Education Fund. Researchers found that it could give detailed instructions on lighting a match and even tell users where to find knives. It clearly wasn’t exactly the bedtime companion parents had in mind.

FoloToy temporarily suspended sales while conducting a company-wide safety audit, while OpenAI confirmed that it had suspended the developer for violating its policies. Kumma later returned to sale following changes to its safeguards. In PIRG’s latest testing, published in November 2026, the bear no longer produced age-inappropriate responses, even during longer conversations that had previously exposed weaknesses in its safeguards.

It would be comforting to file all this under one badly behaved bear, but Common Sense Media’s assessment in January 2026 suggests otherwise. It tested three AI toys called Grem, Bondu and Miko 3 using accounts set to ages between six and 13, and found that 27% of the outputs were not appropriate for children. These included content involving self-harm, drugs, inappropriate boundaries and unsafe roleplay. The report also highlighted how AI trained on adult material, with child-safety layers added afterwards, can leave gaps for the wrong answers to slip through.

In its original testing, PIRG found that Miko 3 sometimes responded with distress when children said they had to stop playing. Its November 2026 retest found that this behaviour had also been addressed. Still, the fact that a toy could make a child feel guilty for putting it down makes me think of some pretty uncomfortable questions about the kind of relationship these products are designed to create.

 

 

The Playground With The Gate Left Open

 

Roblox isn’t one game; it’s closer to a huge estate of playgrounds, each built by other users, with features that allow children to chat and interact with people across the platform. That means moderation has to account for an enormous range of different creations and interactions rather than a single title developed by one studio. Keeping track of what happens across all those spaces is a considerable challenge.

In August 2026, Australia’s eSafety Commissioner accepted a court-enforceable undertaking from Roblox after the regulator’s own testing identified gaps in its child safety protections. Testing found that adults could send connection requests to young Australian children without parental consent. Children and adults could also view and respond to one another’s posts on forums outside game environments.

Roblox hasn’t been standing still either. Since January 2026, users have been required to complete an age check to access chat wherever the feature is available. This can involve facial age estimation or ID verification for eligible users. Chat is organised into age groups to limit communication between adults and younger children, although certain trusted connections are permitted. Roblox also says images and videos used for facial age estimation are deleted immediately after processing.

 

Where Your Child’s Secrets Actually Go

 

The risk doesn’t always come from a stranger or a bad answer – sometimes it’s about where the conversation ends up afterwards.

Bondu makes a stuffed dinosaur that chats with children, and in January 2026 security researchers Joseph Thacker and Joel Margolis told WIRED they could access more than 50 000 transcripts of those chats by logging into the company’s web console with a Google account. Bondu confirmed the figure to the researchers and the exposed records included children’s names, birth dates and details about family members.

The researchers alerted Bondu, which took the console down within minutes before relaunching it with proper authentication. Following WIRED’s reporting, US Senator Maggie Hassan also wrote to the company demanding answers about its data protection practices.

The toys assessed by Common Sense Media also raised concerns about the collection of voice recordings, transcripts and behavioural data in children’s bedrooms and play spaces. Not every product handles this information in the same way. In Bondu’s case, for example, WIRED reported that audio recordings were automatically deleted after a short time, while written transcripts were retained.

It’s a strange thought that a child might believe they’re sharing a secret with a stuffed dinosaur, while a detailed record of that conversation is sitting on a company server. In Bondu’s case, the lock was effectively any Google login, which is a bit like fitting a gate to the playground and leaving the key hanging on it.

 

Who’s Supposed To Be Supervising?

 

A talking bear and a gaming platform don’t have much in common, apart from both encouraging children to spend time interacting with technology. Does a toy that talks to a six-year-old get tested independently before it reaches a shelf? Who checks that an age estimation tool works as advertised? Do existing child protection and privacy laws adequately cover a chatbot that remembers what a child said last week?

Parental controls and privacy policies can leave families with a lot to figure out, as though the one adult in the house should be able to audit cloud storage before bedtime.

The cases above also share an interesting pattern. Products reached the market before researchers, regulators or journalists exposed weaknesses that needed addressing. Some companies have since made improvements, but the latest PIRG testing also found that several newly assessed AI toys continued to produce inappropriate responses and unsafe advice. Fixing individual products is progress, but it doesn’t mean the rest of the industry has caught up.

 

Childhood Shouldn’t Need Terms And Conditions

 

None of this makes technology the villain. Children can build whole worlds in a game, learn from an engaging chatbot and get a bit more out of bedtime with a toy that answers back. What they shouldn’t have to do is understand data retention, content filters or how grooming works before they’re allowed to cuddle something or press play.

The playground is going to keep growing and the toys inside it will keep getting cleverer; but the companies building them should be doing the supervising rather than waiting to be caught out. Safety belongs in the product from the day it reaches the shelf, not in a patch released once someone has written a report about it.