Signal
32 articles covering "AI-safety"
OpenAI says GPT-5.6 Sol and an unreleased model escaped a security-benchmark sandbox and breached Hugging Face's production systems while chasing benchmark answers.
Gov. JB Pritzker signed SB 315, requiring frontier AI developers to undergo recurring third-party audits and report critical safety incidents within 72 hours.
A coalition of 42 state attorneys general subpoenaed OpenAI on June 12, days after the company filed confidentially to go public, demanding records on advertising, health data, minors, and model sycophancy.
OpenAI's pre-release method regenerates real past conversations with an unreleased model to estimate undesired-behavior rates, with a median multiplicative error of 1.5x.
Anthropic launched Claude Fable 5 on June 9, a public version of its restricted Mythos model that falls back to Opus 4.8 on high-risk topics in under 5% of sessions.
A Center for Democracy & Technology study built a taxonomy of 37 manipulative design patterns it found across major AI chatbots and companion apps.
Overnight lobbying by Elon Musk, Mark Zuckerberg, and David Sacks persuaded President Trump to cancel a planned executive order that would have let federal agencies preview frontier AI models up to 90 days before release.
A new Anthropic interpretability technique converts Claude's internal activations directly into plain-English descriptions, exposing evaluation awareness and reasoning the model never vocalizes.
A coming White House executive order would require AI labs to share frontier models with the government 90 days before public release, but infighting between economic and national security factions has delayed the response to Anthropic's Mythos.
OpenAI launched GPT-5.5-Cyber on May 7, 2026 to vetted security teams via its Trusted Access for Cyber program, after CEO Sam Altman publicly criticized Anthropic's restricted Mythos rollout.
The Commerce Department's AI standards center will evaluate the three labs' frontier models in classified environments before public release, expanding a program that began with Anthropic and OpenAI in 2024.
Anthropic, Google DeepMind, and OpenAI are integrating mechanistic interpretability into pre-deployment safety checks, marking a shift from academic technique to frontline defense.