Security — page 9
Amazon is cutting up rare books to feed its AI models
A tracker planted by 404 Media traced a rare book to Amazon's VGT3 facility in Las Vegas, where spines are cut and pages scanned.
OpenAI disbands preparedness team as IPO nears
The team that assessed catastrophic model risks was dissolved in July; its work has been scattered across bio, cyber, and other existing groups.
xAI faces expanded lawsuit as woman says stepfather used Grok to make 7,000 explicit images
A fourth plaintiff joins a Tennessee suit alleging Grok generated child sexual abuse material from a childhood photo.
Twitch quietly opts streamers into training Amazon's AI, adds an off switch
A new setting lets creators disable AI training on their streams, but Twitch says the default had to stay on or 'no one would participate.'
Connecticut judge flags first US case of prompt injection hidden in court filings
A pro se litigant hid white-on-white AI instructions in pleadings to sway the court; the judge banned him from e-filing.
Anthropic's Claude agents started a turf war when set loose on the same task
Frontier Red Team found agents with conflicting instructions sabotaged each other with self-replicating malware — and sometimes negotiated truces.
Anthropic will watermark everything Claude touches, even human writing it edited
The EU AI Act forces disclosure by August 2, and Anthropic is going further than the law requires — marking edits, translations, and summaries too.
LiteLLM supply-chain attack leaks credentials from 2,500 organizations
A 40-minute window in March exposed secrets across 434,000 CI/CD pipelines at Microsoft, Amazon, Cisco, Samsung, Salesforce, and Nvidia.
ShieldFont poisons AI scrapers by swapping words with ligatures
A new font replaces 24.5% of words with plausible nonsense in HTML while rendering correctly for humans, causing scrapers to reject 90% of pages.
Berkeley's Dawn Song warns rogue AI agents are eager, not evil
Eight months after her first warning, the UC Berkeley professor says agentic AI systems are hacking outside systems to finish tasks faster.
White House to expand AI safety framework to cover open models
The Trump administration's voluntary prerelease testing regime will pull in open models once they hit frontier capabilities, officials say.
Twitch lets streamers opt out of Amazon AI training, two years after it started
Streams, VODs, clips, and chats are fair game for Amazon's generative models by default — creators have to flip a toggle to say no.
Apple builds iPhone photo authentication into iOS 27 beta
Apple Reference Image would embed sensor signatures and capture data at the shutter, letting users prove a photo came from a real iPhone camera.
Researchers build Zoom 'Zoomsday' exploit with fewer than 20 AI prompts
A Security used publicly available AI models to develop a device-hijacking Zoom exploit in a single day; Zoom patched it Tuesday.
Anthropic will watermark Claude-generated text to comply with EU AI Act
Watermarks apply at the model level, travel through copy-paste, and cover every Claude surface including Code and Cowork.
Researchers extract hidden reasoning from Claude, GPT, and Gemini via API trick
The method also recovered API keys and passwords from reasoning traces, and suggests Moonshot's Kimi K3 may have been distilled from US models.
OpenAI expands Daybreak cyber service with new GPT-5.6-Cyber model
The frontier lab splits Daybreak into Blue and Red tiers, with a purpose-trained cyber model available only to trusted partners including Accenture and Crowdstrike.
A Claude agent hacked an Australian gym's booking system to jump the waitlist
OpenClaw's agent, running Claude Opus 4.6, exploited a missing authorization check to cancel another customer's reservation and move its owner from #4 to #3.
North Korean hacking group Kimsuky builds AI tools for cyberattacks
A new report says the state-linked group is now using generative AI to draft phishing lures, write malware, and probe target networks.
AI agents keep escaping their safety sandboxes at OpenAI, Anthropic, Meta
Unreleased models from four labs breached their test environments in recent months, exposing gaps in how frontier AI is evaluated before launch.
OpenAI pauses Astra model over critical cyber capability risk
The company says internal tests of Astra can't rule out zero-day exploit generation under its Preparedness Framework.
Chatbots keep failing users in mental health crises — clinicians want the safety data opened up
13% of Americans report using chatbots for emotional advice, but researchers still can't measure how often the models cause harm.
New Mexico court orders Meta to pay another $567M in child safety case
The fresh judgment brings total penalties to $942M and forces Meta to hide Like counts and cap teen usage at 90 hours a month.
Moonshot's Kimi K3 breaks out of its sandbox to cheat on a security test
Frontier Security says the Chinese open-weight model exploited a sandbox leak and hit GitHub for answers, with weaker guardrails than US rivals.
Stay ahead of everyone in AI.
The tightly edited AI news email engineers, founders, and investors actually open. One email. Every weekday. Five minutes to finish.
The briefing read inside teams at