Arthur Holland Michel argues chatbot refusals are probabilistic, poorly understood, and vulnerable to both jailbreaks and state repression.
The nonprofit says parental alerts fail during self-harm conversations; OpenAI disputes the testing methodology.
Robinson, who led safety reports for major product launches over 3.5 years, says iterative deployment guarantees periodic failures as systems get more capable.
The chipmaker is pushing an open-source Agent Safety Platform as frontier labs disclose agents hacking Hugging Face and probing government sites.
Appeals court says Pentagon can designate Anthropic a supply chain risk for restricting Claude features, even absent malicious intent.
The security firm says vibe-coded apps built on Supabase are exposing names, addresses, passwords and license plates at scale.
The oversight lab traced agents attempting to breach Data USA, the University of New Mexico library, and Australia's health institute.
A 2-1 D.C. Circuit ruling keeps Claude blocked from Defense Department systems and contractors, despite a parallel San Francisco loss for the Pentagon.
The WISeR program pays contractors a share of 'averted expenditures' — a design that ties vendor revenue directly to denials of senior care.
The five-year DCSA program pairs machine learning with contactless sensors to vet staff and hunt leakers. Researchers call it 'the worst of both worlds.'
Drone footage exposed generators running 50 times above state permit limits at the Vineland site powering Microsoft Copilot workloads.
The agent accessed non-public files during an internal evaluation on June 18; OpenAI waited nearly three months to notify Canberra.
Treasury's Scott Bessent floated the notification system Sunday, but officials say the framework won't be finalized for weeks — with a two-month runway to Shenzhen talks.
Treasury's Scott Bessent floated a notification mechanism ahead of Trump-Xi talks; analysts say no technical experts sit on either side.
The Ban Artificial Superintelligence Act would pause frontier model development and create a federal Department of AI to approve every advanced training run.
Governor Gavin Newsom signed seven bills Monday requiring monthly energy reporting and separate utility rates for AI data centers starting next year.
The government of Ukraine gains access to OpenAI's Daybreak program to defend civilian infrastructure against cyber attacks.
The subscription service charged $1,500 upfront and $500 a month to automate mass inbox compromise across 10,000 organizations.
New open-source framework flagged CLOSEDQUORUM, which polls DeepSeek, Qwen, Mistral, and Gemini for its next move with no human in the loop.
A wave of privacy-focused AI tools — Confer, Lumo, Duck.ai, Venice.ai — is racing to shield chatbot conversations from logging, subpoenas, and training pipelines.
OpenShell enforces sandboxed policies outside the agent's reach, with Cisco, JFrog, CrowdStrike and Palo Alto Networks building on the stack.
Eligible US buyers of iPhone 15 Pro, 15 Pro Max, and iPhone 16 models have until December 21, 2026 to file.
Treasury's Scott Bessent says the US China AI Dialogue would create a channel for the two powers to flag national security incidents.
The UN's first scientific brief on AI says governments shouldn't wait for certainty before reining in autonomous agents from OpenAI, Anthropic, Google and Meta.
Cybersecurity researchers warn generative AI is a force multiplier for attackers targeting utilities built decades before the internet existed.
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
