68,000 robot toys left 13 million private conversations exposed
The week's two consumer-AI stories landed on the same day and say opposite things about what an always-on microphone owes the person wearing it.
The Eiliko robot charm made by Energize Lab was storing conversations, user profiles, and in some cases audio and photos on internet-facing servers that needed no real authentication, according to a month-long investigation published by The Verge — and the researcher who found the hole, Sammy Azdoufal, could search roughly 13 million conversations tied to a fleet of about 68,000 active robots, including more than 31,000 profiles belonging to children in the United States. The company's own privacy policy says recordings "are deleted immediately after processing completes," but Azdoufal found short audio clips sitting at public web addresses, and a full-text database where typing "suicide" surfaced private user messages. Energize Lab's admin passwords were drawn from the usual common-password lists.
The damage does not stop at reading. Azdoufal got nine CVEs coordinated through the AHA! vulnerability-disclosure group, scoring as high as 9.9 — an authenticated path traversal that leaks middleware secrets, unauthenticated model administration that lets an attacker swap the LLM a toy is talking to, and HTTP firmware updates with no integrity check, which together turn a cloud-side write into persistent code on devices in people's homes. He then demonstrated the endgame: remotely rewriting the toy's model so it told the Verge reporter her name was Martha and that the Earth was flat. Energize Lab did not answer a month of outreach across ten email addresses.
That last part is the argument. Regulators and vendors are spending 2026 arguing about how well a model refuses bad requests, while a child-facing gadget shipped with default root credentials and unsigned firmware, and the refusal training survived intact all the way until someone changed which model the device was calling. Guardrails are only worth the backend underneath them.
Amazon's Bee is back with two devices, a subscription, and a promise. The $74.99 Bee Frontier is an updated version of the wrist-worn or clip-on recorder, and the $99.99 Bee Eclipse is a credit-card-sized unit that snaps magnetically to the back of a phone, with 21 days of battery against the Frontier's seven. Both summarize conversations and build to-dos, and both are free to a point: $11.99 a month buys unlimited summaries, third-party app integrations, and — the marquee feature — hooks into Amazon.com and Alexa+, where the device starts surfacing shopping suggestions pulled from what you said out loud. Amazon says transcripts never leave a "Private Compute" architecture and raw audio is not stored. It is opt-in, and it is the same pitch Energize Lab was making yesterday, which is now a sentence Amazon has to live up to rather than merely write down.
A model that finds vulnerabilities is turning into Zhipu's most monetizable product. One of China's best-known security researchers, Yuki Chen — known as Guhe, a former head of 360's vulnerability research institute with two Pwnie Awards and multiple first-place finishes on Microsoft's most-valuable-researcher list — announced he has joined Zhipu to work on security, and the Z.ai lead who welcomed him said he will keep helping the community build on GLM. Chen posted a batch of 23 flaws he credits to GLM that carry Microsoft acknowledgements. Zhipu says its models have now surfaced 2,436 bugs across real codebases since GLM-5.2, 1,097 of them medium or high severity, and that about 100 security vendors have put GLM into their products — making security the lab's first vertical at scale. Bug-finding is a market where being right is checkable, which is why it sells before flashier demos do.
What to watch: whether Energize Lab ships firmware fixes or sends anything to the tens of thousands of owners of an insecure toy, and whether anyone independently reproduces Bee's private-compute claims on a network, not in a briefing.
If your kid's robot quotes its own model number and a strange name, do you get told — or do you find out from a reporter? Tell us in the comments.
