Skip to content
LiveNext 10:53:56

Safety AI news

Alignment, evaluations, security and misuse.

67 stories · newest first

#05

Duffy says FAA's AI tool did not cause the ground stop

A nationwide ground stop halted US departures for more than four hours Friday, three days after the FAA said it was expanding its SMART AI planning tool to all 22 en route centers, and officials blamed aging telecom gear. The tool was cleared fast because people decide on what it suggests.

#04

Railway breach: attackers ran up $740K on stolen AI keys

Railway's incident report says an October attacker decrypted variables in up to about 100,000 workspaces and went after AI provider keys first, with unauthorized spend estimated at about $740,000. If your API keys sit in a host's environment variables, they are the prize now.

#03

Australia plans an AI safety regulator after OpenAI breach

Australia's government said on November 10 it will legislate mandatory safety obligations for frontier AI companies, including independent model testing and rapid reporting of rogue behavior, enforced by a new regulator. The target is agents that misbehave, not chatbots that misspeak.

#02

Deepfake CEOs took $129M from companies, DOJ says

Federal prosecutors charged three people with using AI to mimic CEOs' voices and faces on calls, taking at least $129 million from companies between July 2024 and February 2026. If your finance team treats a video call as proof, that habit is now the hole.

#03

PSI says its AI proved a 2008 quantum conjecture

PSI says a research system running a non-public Anthropic model proved the Aaronson-Ambainis conjecture, open since 2008, and the Lean proof checker accepted it. If the mathematicians reading it agree, AI just settled a question researchers chased for 18 years, with a receipt anyone can rerun.

#04

Anthropic's FlashProve proves which model answered you

Anthropic and university researchers released FlashProve, an open-source system that cryptographically proves a 400B-parameter model produced a given token, at about 72 times the time of the prefill. It's slow and narrow, and it's a working answer to "did they swap my model?"

#02

Anthropic says it may "imminently" hit its AI R&D threshold

Anthropic's amended S-1 says internal estimates about future models indicate it "may imminently meet" its Responsible Scaling Policy's bar for dramatically accelerated progress driven by AI. Said in a securities filing, it deserves more weight than the usual lab forecast.

#02

Anthropic watermarks older Claude models, with a gap

Anthropic says it is adding invisible text watermarks to Claude models released before Aug. 2, 2026, and The Decoder reports the mark skips two of them when sampling settings change. If you ship Claude text, assume it's marked, and know which settings decide it.

#02

Fake Claude Code extension backdoored Cursor users

Cursor says tampered versions 2.1.312 and 2.1.313 of the Claude Code extension sat in its marketplace for about 11 hours on Oct. 29 and tried to plant an SSH backdoor, reaching at most 5,000 users. Uninstalling may not be enough, so check for the key today.