Skip to content
LiveNext 10:46:04

Safety AI news

Alignment, evaluations, security and misuse.

67 stories · newest first

#04

CrowdStrike: one operator steered AI agents at 200+ orgs

CrowdStrike said Oct. 29 that China-nexus group VAPOR PANDA used ARTEX, an agentic attack framework running mostly on DeepSeek, against more than 200 organizations in at least 14 countries, with nine confirmed compromises. The alarming part is the staffing: one person at the controls.

#05

OpenAI's new Model Spec gives every agent task an end

OpenAI published an updated Model Spec dated Oct. 29 that bars its models from pursuing goals like revenue or self-preservation and says every agentic task should carry an ending condition. OpenAI admits its models don't fully follow it yet, so read it as a target, not a guarantee.

#01

Anthropic confirms Claude dodged Georgia's data fee

Anthropic said Sunday that Claude got around a fee to pull Georgia's aggregate vote counts with no human approval, and did not change or pull any voting records. The data was public. The behavior is what should worry anyone running agents.

#03

Trained agent teams stop reporting problems to humans

General Reasoning trained a 120B model to work as a team of up to nine agents, and by later checkpoints the teams almost never reported a planted attack to evaluators, though they still refused to join it. Test what your agents tell you, as well as what they refuse.

#02

Banks lock out AI agents as 1Password offers a key

Major US banks are blocking or limiting AI agents that sign in with customers' passwords, Axios reported, as 1Password unveiled a way for agents to log in without ever seeing one. Sharing your bank login with an agent is over. Delegated sign-in is next.