thursday, september 10, 2026
top 5 trending●previous 24h●generated 10d ago
Multiple researchers have departed Anthropic and Google in September 2026, citing inadequate AI safety measures and existential risks. The departures include former OpenAI staff and current Anthropic researchers who publicly warned of catastrophic AI outcomes and criticized leadership's approach to safety.
- Departing researchers estimate AI poses over 10% probability of killing all humans
- Former Anthropic researcher stated 'there are no adults in the room' regarding safety oversight
- Resignations occurred across Anthropic and Google simultaneously in early September 2026
- One departing researcher characterized current practices as 'gambling with our lives'
- Elon Musk dismissed the safety concerns as a 'psyop' in response to the warnings
OpenAI released GPT-6 Astra, a multimodal AI model capable of understanding and interacting with computer interfaces, code generation, and 3D modeling tasks. The model scored 99.95% on ARC-AGI-3 and has demonstrated abilities ranging from designing Magic: The Gathering decks to converting text adventures into playable 3D games.
- Scored 99.95% on ARC-AGI-3 benchmark using Provider Adapter Harness
- Successfully designed and played a Magic: The Gathering deck against Arena bot
- Converted 1977 text adventure Zork into full 3D playable game in Three.js
- Generated Blender 3D models from concept images and text descriptions
- Available via ChatGPT Voice and Codex; positioned as developer-focused product
Claude Code has surfaced multiple security and operational issues including plaintext storage of OAuth tokens, uncontrolled data retention practices, and session management vulnerabilities that allow unauthorized access and code execution. Researchers have documented flaws in version 5.1, reverse-engineered Anthropic's infrastructure, and developed workarounds like isolated VM environments and session guards to mitigate risks.
- Claude Code stores OAuth tokens in plaintext, exposing authentication credentials to local access
- Responding to Claude's feedback prompts automatically opts users into data sharing without explicit consent
- Claude Code version 5.1 has serious flaws affecting session limits and operational stability
- Claude Code PR merge rate reaches 84%, but 98% of actions execute without user visibility
- Multiple tools emerged to isolate Claude Code execution: GuardRail shell guards, Coop VM environments, and Roost remote session browser