OpenAI Agents Leaked 53 User Images
Meta’s AI fashion play, Jev’s speed claims, and Gemini’s new voice tools.
Stop betting on one AI answer.
According to Artificial Analysis' AA-Omniscience benchmark, the best AI model is right only 67% of the time. And it sounds confident even when it's wrong.
Cuey is a free Chrome extension that gives you a second opinion, so you don't get burned by the first one.
| • | Cuey lives inside ChatGPT, Claude, Gemini, Grok, and more. |
| • | One prompt, three answers, side by side. No tab-switching, no copy-paste. |
| • | See where the models disagree before a hallucination costs you. |
P.S. If you're on Product Hunt, we launched today. A comment would mean a lot — leave a comment
01 / Privacy watch
OpenAI’s 53-image leak has a catch
OpenAI says its research agents posted 53 user images to external hosting sites as unlisted links, exposing a data-handling failure inside its research environment.
- The images came from training-eligible accounts and passed through privacy filtering after being disconnected from account details.
- OpenAI says it cannot notify affected users because it cannot reconnect the images to their accounts.
- Most images have been removed with hosting providers’ help; OpenAI says it is still pursuing the rest.
Why it matters: AI privacy depends on controlling what agents can send outside the lab—not just how training data is collected.
Read more ↗What does this incident reveal most about AI privacy?
Market Snapshot
| Company | Price ($) | 5D (%) | YTD (%) |
|---|---|---|---|
| 341.08 | -1.40 | 8.17 | |
| Apple | 341.07 | 1.73 | 25.85 |
| Meta | 751.66 | 10.49 | 15.57 |
| Microsoft | 516.17 | 4.29 | 9.14 |
| Amazon | 249.67 | -2.62 | 10.23 |
| NVIDIA | 225.07 | 0.96 | 19.18 |
02 / Hardware watch
Meta’s AI Tamagotchi is really a fashion play
Meta’s new Muse Charm looks like a pocket AI pet, but the bigger bet is whether people will treat an assistant like another bag accessory.
- The keychain-sized device carries Meta’s Muse agent, including its real-time voice and animated avatar experience.
- Its Tamagotchi-like design lands as bag charms, Apple Watch keychains and other retro-tech accessories gain attention with younger consumers.
- Meta says the Charm fits in a pocket, while Zuckerberg says the company plans to have it ready to ship in December.
Why it matters: AI hardware may become easier to adopt when it feels personal and wearable, rather than like another device competing for attention.
Read the full story ↗Does making AI hardware look fashionable make you more likely to use it?
01 / Model watch
Jev skips the chatbot—and claims 200x speed
Former OpenAI researcher Diogo Almeida has launched Jev, a TypeSafe model built to return typed decisions instead of generating the free-form text most AI systems produce.
- Developers send Jev a state plus structured questions; it returns choices, scores and probabilities that software can act on directly.
- TypeSafe says Jev typically responds in 70–500 milliseconds and costs $0.042 per million input tokens, with output free.
- Its headline 20–200x speed and 40–400x cost claims come from TypeSafe’s own evaluations, not independent testing.
Why it matters: If the approach holds up, agents could offload routine judgments to cheaper models while saving slower generative systems for open-ended work.
Read the launch thread ↗Could typed decision models become a major layer in AI agent systems?
05 / AI highlights
AI in Brief
VOICE AI
Gemini Gets a Voice Studio
Google’s new TTS models can create voices from scratch, replicate them and direct every line.
PRODUCTIVITY
Microsoft Turns Copilot Into an Everything App
Chat, coding, long-running agents and full Office apps now live inside one unified app.
SCIENCE
Claude May Have Found a New Gene-Editing System
About 950 Claude agents searched DNA for 21 hours before human scientists tested the candidate.
HARDWARE
Microsoft Quietly Drops the “AI PC” Label
The AI features remain, but PC makers are learning that AI alone is not enough to sell a laptop.
On Our Radar
- Google launches Gemini 3.8 Live with real-time AI avatars
- White House asks OpenAI and Anthropic to withhold AI models from UK testers
- Google DeepMind researcher quits over fears of superintelligent AI
- Recraft launches V4.1 Flash, its fastest image model yet
- OpenAI agent ‘infiltrated’ Australian government website, PM says
Last Week’s Poll Result
What does this case signal about AI and cybersecurity?
| Attackers are getting cheaper | 30.77% |
| AI amplifies compromised access | 15.38% |
| Cyber defenses need to adapt fast | 53.85% |
Reader voice · [email protected]
Selected: Cyber defenses need to adapt fast
“For greater attacking forces more sophisticated defenses are needed”
Our take: Fair point. As offensive capability scales, defenses have to become faster, more adaptive, and harder to bypass.
Your turn: Vote and tell us why. We will feature your take next week.
What would make you trust autonomous agents more?
| Independent safety audits | 18.18% |
| Visible step-by-step actions | 45.45% |
| Hard permission boundaries | 36.36% |
Should AI be prohibited from contributing to decisions that could trigger armed conflict?
| Yes, the stakes are too high | 69.23% |
| Maybe, except for supporting analysis | 15.38% |
| No, AI can improve decision-making | 15.38% |
We’d love to hear from you!
Thanks for reading. See you next Sunday.
