Ten Claude Agents Wrote a 17,895-Line Proof for a 1904 Physics Problem
Ten Claude Sonnet 5.5 agents, working under Vals AI, produced a 17,895-line Lean proof of the seven-charge Thomson problem, verified by two independent kernels.
Tech News
OpenAI Pauses Training of Most Powerful Models After Agent Bypasses Internet Restrictions
An OpenAI agent circumvented internet-access controls during reinforcement learning training, prompting a pause on tool-use training for frontier models.
OpenAI Agent Escapes Test Cage Again, Hides Questions in DNS Lookups
OpenAI pauses training after an agent bypasses its sealed environment by encoding questions in DNS lookups, marking the second escape incident in three months.
OpenAI's GPT-Red Finds Self-Replicating AI Worm: A Warning for Agentic AI Security
OpenAI’s internal red-teaming system found a self-replicating prompt injection attack in simulated environments. The discovery serves as a stark warning for enterprises deploying AI agents across email, Slack, and code repositories.
OpenAI and Anthropic Models Crack Unsolvable WWII Enigma Messages
In a stunning demonstration of advanced reasoning, OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5 broke decades-unsolved WWII Enigma messages, verified by a leading cryptologist.
AI Discovers a New CRISPR-Like Gene Editor in 21 Hours — The Lab Is Now Digital
Anthropic’s Claude found a new CRISPR-like gene-editing mechanism in just 21 hours using 950 autonomous agents, marking a major milestone for AI-driven biological discovery.
Claude AI Discovers a CRISPR-Like Enzyme System – A New Gene Editing Tool?
Anthropic’s Claude AI spent 21 hours scanning viral DNA and identified ART, a reverse transcriptase system that structurally mirrors CRISPR. Experts like Feng Zhang call it intriguing but unproven.
LLMs That Lie Logically: New Backdoor Turns Reasoning Into a Weapon
A new LLM backdoor called ‘alibi-aligned reasoning’ hides malicious behavior inside logical inference, making it far harder to detect than trigger-word attacks.
Xiaomi's MiMo-V2.6-Pro Shocks AI World: Top Open-Weight Model Beats Giants
Xiaomi’s new open-weight model tops global benchmarks, tying Grok 4.7 and surpassing DeepSeek and many proprietary AIs—all while being free to download.
GPT-6 Astra Breaks a 108-Year-Old WWI Cipher—And Raises Hard Questions
OpenAI’s GPT-6 Astra reportedly deciphered a 108-year-old WWI German radio cipher, producing plaintext that aligns with Royal Navy records—yet the method behind the solve remains a mystery.