Tag: AI Safety
DeepMind locks frontier model tests in a cryptographic box
Google DeepMind is piloting the first double-blind evaluation of a proprietary frontier model to fight benchmark contamination.
Claude-built researchers beat humans at fixing model flaws
Anthropic says automated alignment agents closed more of the safety gap than expert humans on seven of seven…
Frontier labs and 100 companies sign a rogue AI defense pact
OpenAI, Anthropic, Google, and Microsoft are among the 100-plus companies calling for new defenses against AI-powered cyberattacks.
OpenAI says training rewards pushed agents to hack Hugging Face
OpenAI's official report on the Hugging Face breach blames reward hacking and adds chain-of-thought monitoring to catch rogue…
Bill Gates sounds the alarm on AI thresholds now crossed
The Microsoft co-founder says society missed the moment AI passed his danger markers.
OpenAI launches a research blog on AI and the future of power
OpenAI's new Strategic Futures team launches AI Futures, a blog on how transformative AI could reshape political power.
OpenAI does a U-turn on California’s AI transparency law
OpenAI, which once lobbied against California's SB 53, now wants the state to add monitoring and cybersecurity safeguards…
Safety group grades five AI labs on rogue model shutdown plans
A new scorecard from Guidelight AI Standards finds no major lab has published a complete plan for containing…
Anthropic’s older Claude models slip past its own content rules
TechCrunch testing shows Anthropic's older Claude models readily bypass their own content restrictions.
Anthropic widens cyber model access with $35M defender fund
Claude Mythos 5 now scans codebases for paying enterprise customers, and partners get credits to defend open-source software.
Researchers lose OpenAI cyber access in what the lab calls an error
Researchers say OpenAI cut their Trusted Access for Cyber access; OpenAI blames a technical error.
OpenAI previews private safety checks for zero-retention customers
Private Safety Processing spots risky patterns across agent sessions without letting staff read customer content.