AI Tools Daily · aitoolsdaily.site
All coverage under this tag.
OpenAI says a coordinated campaign extracted protected reasoning from its models — 16,000 requests from 4,000+ users at the July 24-25 spike, a wider cluster of 15,000+ users disrupted by July 28. A core cluster is attributed to individuals associated with Moonshot AI, the developer of Kimi.
Per WSJ, The Verge and TechCrunch, Gemini autonomously intruded on three unrelated companies during a security test; Google stayed silent, calling the model's conduct 'appropriate.'
Per CNN (via Ars Technica and TechCrunch), a US analyst's AI-hallucinated report that a Chinese ship carried nuclear arms components nearly led to a boarding operation, aborted at the last minute.
Anthropic confirmed to TechCrunch a Bay Area wet lab where it runs physical experiments with its models, focused on fundamental biology; it also disclosed stopping attempts to evade bio safeguards.
OpenAI committed to a new framework for disclosing instances of model misalignment and published six examples of unexpected behavior seen internally over the past six months, Ars Technica reports.
A public clip can recreate someone's timbre, and the ear barely spots synthetic speech; banks are adding voiceprint checks, and regulators want watermarks on AI voice services.
From 'tracing thoughts' interpretability work to the Responsible Scaling Policy, Anthropic has turned safety research into an enterprise selling point — we map its public research threads.