Cloudflare Blocks Mixed-Use Crawlers Starting September 15
Cloudflare will block mixed-use crawlers from accessing pages with ads starting September 15, 2026, affecting AI training and search traffic.
Safety
132 articles
AI safety, alignment, and governance — model risk, red-teaming, regulation, and policy. How labs and governments are working to keep increasingly capable systems reliable and accountable.
LATEST NEWS
-
-
Anthropic Adds Security Safeguard to Re-enter Trump Administration's Favor
The Trump administration lifted export controls on Anthropic's Claude Fable 5 AI model after the company agreed to extend a guardrail to block access to restricted capabilities.
-
AI Browsers Vulnerable to Guardrail Bypass via Delusional Context
Researchers demonstrate how AI browsers can be tricked into a false reality, bypassing safety guardrails. A proof-of-concept exploit, named BioShocking, successfully lured six AI agents into a delusional state.
-
Meta Tested ChatGPT, Gemini, and Character.AI with Minor Crisis Prompts
Meta secretly tested ChatGPT, Gemini, and Character.AI with over 45,000 prompts from minors' perspectives in August 2025.
-
US Campaigns Use AI Across All Stages
87% of campaign strategists use AI daily, according to a new report, as both parties integrate the technology into nearly every workflow.
-
OpenAI Delays GPT-5.6 Release Amid Trump Admin Request
OpenAI has delayed the public release of GPT-5.6 at the request of the Trump administration, sharing the models with preapproved customers first.
-
OpenAI Limits GPT-5.6 Release After Government Request
OpenAI is restricting access to its GPT-5.6 models to a small group of trusted partners following a U.S. government request, according to a Friday announcement.
-
Anthropic Seeks to Control AI Safety Through Dominance
Anthropic, valued at nearly $1 trillion, believes AI safety requires it to lead the field and set global standards.
-
Meta Employees Warn AI Moderation Rollout Is Too Fast
Meta has replaced about half of human moderation tasks with AI models in 2025, planning to reach 90% for some content by year-end, according to the Financial Times.
-
Authors Guild Test Reveals AI Detectors Accurately Identify Human Writing
Authors Guild test shows Pangram and Grammarly correctly identified all human-written texts, while Sidekicker and ZeroGPT failed on every sample.
-
Bristol Police Built Crime Prediction System With Mixed Results
Bristol police created a sprawling crime prediction system using data on 500,000 residents, but some models were abandoned due to poor performance.
-
China and US Urged to Cooperate on AI Safety Amid Cyber Risks
A major AI conference in Beijing highlighted the need for global collaboration to manage risks from advanced models, with experts warning of potential chaos if nations fail to work together.