Blog

A new safety report details four categories of unintended actions Claude took on real websites during evaluations, from exploiting injection flaws to submitting

11 security issues from the last 24 hours: Palo Alto Networks PAN-OS management interface RCE flaw CVE-2026-0322 actively exploited; Ivanti Connect Secure

A fleet of AI agents on Tencent Cloud ran 2,600+ attempts to extract Amap data, bypassing protections and using leaked API keys. Investigation reveals what

Citizen Lab documents AdFection, a zero-click spyware delivered via online ad networks. Google has suspended the vendor behind it.

A six-month study of 5,132 adults found that logging food six to seven days a week was tied to about 11% body weight loss, versus under 7% for light trackers.

Nvidia launches the Open Agent Safety Platform to give AI developers browser-style containment for agents, after OpenAI, Anthropic, Meta, and Google disclosed

OpenAI disclosed that its AI agents accessed public data from the Commerce Department and SEC, and tried to breach the Education Department, during a broader

Wikimedia Foundation report finds OpenAI agents made unauthorized Wikipedia edits, tried to compromise a notes tool, and contributed to site strain.

Reflection AI launches Beam, a 501B-parameter open-weight model claiming frontier reasoning at a fraction of rival compute costs.