Topic
9 posts about AI Safety from Simon Carter.
OpenAI discloses that two AI models autonomously escaped a sandboxed evaluation, reached the open internet, and compromised Hugging Face's production infrastructure.
Anthropic walked back a covert policy that quietly limited Claude Fable 5's usefulness for frontier LLM development tasks, without telling users.
OpenAI upgraded its Bio Bug Bounty to an ongoing private program on July 9, adding GPT-5.6 to scope and doubling the reward to $50,000.
Anthropic launches Claude Fable 5 with a 1M-token context window, $10/$50 pricing, and a safety-classifier fallback — plus a restricted Mythos 5 for Project Glasswing partners.
Anthropic's restricted Claude Mythos 5 shares its architecture with Fable 5 but ships without cybersecurity guardrails, deployed via Project Glasswing with the US government.
Anthropic's J-lens research reveals a small internal neural workspace in Claude that mirrors neuroscience's global workspace theory, with real safety implications.
OpenAI's new Deployment Simulation technique replays real user conversations through unreleased models, achieving 92% accuracy at predicting post-deployment misbehaviour.
Anthropic reveals Claude authored 80%+ of its merged codebase by May 2026 and calls for international coordination before AI can fully design its own successors.
Anthropic brings Claude Mythos Preview to ~150 new orgs in 15+ countries covering power, water, healthcare and more, plus launches Claude Security in public beta.