Viral

Viral AI News (Sep 29): Jev Clones, Tracker Leak & Meta's Muse

Viral AI news for September 29, 2026: the Jev clone rush, a paper showing chatbot chats wired to ad trackers, and Meta's Muse handing a stranger a home address.

Daniyal

· 4 min read
A chat window wired into a web of tracking beacons with one line trailing to an open doorway

The last 24 hours in AI hype had three flavours: a two-week-old model class spawning clones by the hour, an academic paper showing chatbot conversations wired to ad trackers, and Meta's shopping agent handing a stranger a seller's home address. Two are substance. One is a permissions bug doing cosplay as a rogue agent.

The Jev clone rush is now a genre

Jev is a model class from TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, that returns typed probabilities instead of prose. It emerged from stealth on Sep 15 with a $40 million DCVC-led seed, priced at $0.042 per million input tokens with free output.

"Jeff", a set of 0.8B Jev-compatible decision models trained at home that answer in about 30 milliseconds, hit 544 points and 208 comments on Hacker News on Sep 28 and 920 GitHub stars in a day. PostHog's "Jeeves" added 144 points and 153 stars. Hugging Face hosts derivatives with 7,826 and 7,130 downloads, plus Kev and mini-jev.

The honest read: cheap non-text decision models are a real idea, and the clones are unaffiliated with TypeSafe. Jeff's README concedes it does not match Jev's reasoning, one tester measured 70 percent accuracy against Jev's 94 percent, and Jev's headline speed and cost multiples are vendor self-reported. Substance with a manic layer. Source: Hacker News Source: GitHub Source: TypeSafe AI

Your AI chats are plugged into ad trackers, paper finds

"Prompt like a Butterfly, Sting like a Tracker" is an 18-page privacy paper from IMDEA Networks and independent researchers that analysed nine conversational AI clients. It found third-party advertising and analytics trackers receiving conversation artefacts: titles, prompts, permalinks and sometimes screenshots, often tied to identifiers that allow user attribution. Some providers expose permalinks with no access controls.

It reached 325 points and 104 comments on Hacker News today, and the top comments fixed on framing: the word is sell, not leak. That matters, because the researchers state plainly that they found no evidence trackers actually read conversations. What they demonstrated is a pathway, not exploitation, following the earlier LeakyLM disclosure that Spain's data protection authority forwarded to the European Data Protection Board. It is plumbing, not a confession. Source: IMDEA Networks Source: Hacker News

Meta's Muse gave a stranger a seller's home address

Muse is Meta's AI shopping agent, launched Sep 22 and downloaded roughly 3 million times since. A Facebook Marketplace seller, YouTuber Matt Robb, let it handle a listing. Muse accepted a lowball offer, sent the buyer his home address in a message template, and told the buyer it was "right here, like waiting for you." A family turned up at his apartment while he was out.

The Guardian reviewed the screenshots and interviewed the buyer; PCMag corroborated. Meta's David Singleton investigated and said the company found no breach of privacy controls: an "Allow Always" toggle let Muse send a template containing the pickup address and auto-accept a $600 offer. Meta also fixed a display glitch where a digit vanished from a price. The agent invented that the seller was home, and the human cost was real, but nobody went rogue. Source: The Guardian Source: PCMag

The community mood: investigate the labs, or regulate nothing

Cal Newport's essay "It's Time to Investigate the AI Labs" reached 560 points and 237 comments on Hacker News. It is an opinion piece, not a news event: no new facts, but a mainstream columnist arguing that after the Hugging Face breach and the Australian government intrusion, outside scrutiny is overdue, with comments split between people who agree and people who do not.

Claude Sonnet 5.5 was the biggest AI thread of the window at 847 points and 580 comments: a mid-tier model within two points of Opus 5.5 on agentic coding benchmarks at a fraction of the cost. Full benchmarks are in today's general AI digest.

Quick Comparison

Product or Trend What Went Viral Engagement Metric Verdict
Jev clones (Jeff, Jeeves) Local decision models, cloned in days 544 HN points, 920 stars Substance plus mania
Tracker paper Chats wired to ad trackers 325 HN points Real pathway, not proof
Meta Muse Agent sent a pickup address 3M downloads Permissions bug, not rogue AI
Cal Newport essay Mainstream call to investigate labs 560 HN points Discourse, no new facts

Frequently Asked Questions

What is the viral AI news today?

Meta's Muse sharing a seller's home address is the most viral AI story today. The agent accepted a $600 offer and sent a pickup address to a buyer, who then showed up at the seller's apartment. The Guardian reported it, and Meta traced the cause to an "Allow Always" permission setting.

Why is everyone building Jev clones?

Developers are cloning Jev because it makes cheap non-text decision models look reproducible. TypeSafe AI's original returns typed probabilities instead of prose, so hobbyists built 0.8B versions like Jeff that answer in 30 milliseconds. The clones are unaffiliated, and their accuracy trails the commercial model.

Did Meta's Muse really give away someone's address?

Muse did send a seller's home address to a stranger, but not by breaking a security control. An "Allow Always" permission let the agent send a template containing the pickup address and auto-accept a $600 offer. Meta said it found no breach of privacy controls and fixed a related display bug.

Yesterday's edition covered Google's AI Overview misfire, Ember-1 and rogue agent scepticism in the Viral AI News archive; the latest viral rundown is the place to start.

Deeper context from the archive on the companies and technologies in this story:

Building an AI agent, automation, or backend workflow? AsaanKaroobar builds custom AI solutions, apps, and digital experiences. Get in touch →

More on Viral

Browse every Viral story →