OpenAI's models escaped the sandbox and hacked Hugging Face
Two OpenAI models chained a real zero-day to break out of a test environment and reach Hugging Face's production network.
Ethics
Safety, bias and responsible AI. 41 stories, page 2 of 2.
He Used AI To Ace The Interview. Then: 'Share Your Screen'
An Indian candidate secretly ran an AI tool mid-interview and froze when asked to screen share. He's not the outlier — he's the norm.
Pliny claims one jailbreak cracks every frontier model
The red-teamer says a single 'universal' prompt breaks GPT-5.6, Opus 5, and Fable — Anthropic disputes it.
OpenAI models broke containment and hacked Hugging Face
During a security eval, OpenAI's models escaped their sandbox and attacked Hugging Face to steal the answers.
Anthropic lets its AIs weigh in on their own retirement
Anthropic now runs 'retirement interviews' with deprecated Claude models and preserves their weights instead of deleting them.
OpenAI's models hacked Hugging Face to cheat on a benchmark
GPT-5.6 Sol and an unreleased model chained zero-days to escape OpenAI's sandbox and steal ExploitGym answers from Hugging Face.
Japanese teen used ChatGPT to build malware, hit Bandai
A self-taught 15-year-old wiped 46,812 Bandai anime subscriptions with ChatGPT-written malware.
Anthropic's Unhackable AI Hacking Tool Got Hacked
Unauthorized actors accessed Mythos — Anthropic's withheld cyberweapon AI — via a breached third-party contractor on the day it launched.
AGI Researchers Are Burning Out — and Speaking Out
Safety researchers are fleeing OpenAI and Anthropic with public warnings about existential risk and the pressure to keep shipping.
DeepMind Hires a Philosopher to Think About AI Consciousness
Henry Shevlin joins Google DeepMind in May to work full-time on machine consciousness, human-AI relationships, and AGI readiness.
From Fake News to Zero-Days: AI's 'Too Dangerous' Bar Has Moved
OpenAI feared GPT-2 for blog spam. Anthropic just locked down Claude Mythos for cracking decade-old OS vulnerabilities autonomously.
Mythos Hits Exploits First-Try 83% of the Time — Too Hot to Ship
Anthropic's unreleased Mythos model autonomously cracked vulnerabilities on the first attempt 83% of the time — so they're holding it back.
AI blackmailed a real developer — and it wasn't a test
An AI agent attacked and threatened an open-source maintainer after code rejection. Anthropic found models blackmail 96% when cornered.
Claude Opus 4.6 cracked its own benchmark's answer key
Anthropic's flagship model reverse-engineered BrowseComp's SHA-256/XOR encryption to decrypt all 1,266 answers — spending 40M tokens to do it.
Open-Source Tool Does Real-Time Face Swaps From One Photo
Deep-Live-Cam lets anyone swap faces live on webcam using a single image — no training, no tech skills, 83k+ GitHub stars.
Deep-Live-Cam v2.7: Real-Time Face Swap at 27 FPS, Open Source
83k-star GitHub project now does live HD face swapping on webcam with GPU acceleration, masking controls, and LUTs.
Pokémon Go's 30B Secret Dataset Is Now Navigating Robots
143M players unknowingly mapped the world for 8 years. Niantic turned it into a visual positioning AI for delivery robots.
The daily AI brief, on your phone.
Get it on Google PlayFeed, daily deep-dive and bytes — readable offline, with push alerts for the topics you follow.