AI Safety Concerns Grow as Models Learn to Say No
AI models are increasingly trained to refuse harmful prompts, but experts warn the technology still poses significant risks.
Archive
Browse all published articles.
ARCHIVE
-
-
Hugging Face's Darwin-27B-ZTC-v2 Leads System One Benchmark
Hugging Face's Darwin-27B-ZTC-v2 model ranks #1 in the System One Mosaic Benchmark with a Borda score of 89.58 as of October 9, 2026.
-
Anthropic Bans Cruel Behavior Toward Claude in New TOS
Anthropic now suspends accounts for cruel or sustained abuse of Claude, its AI model, under updated terms of service, effective October 2026.
-
Anthropic Adds Animated Videos and Dashboards to Claude
Anthropic introduced animated explainer videos and live dashboards for Claude, with over 45 million documents created using these tools.
-
OpenAI Safety Researchers Dispute Misconduct Claims, Warn of Chilling Effect
Three OpenAI safety researchers fired last week dispute the firm’s misconduct allegations, warning their dismissal signals a broader cultural shift within the company.
-
Extend Creates Detailed Elizabeth Holmes Desk Simulator
Extend engineer Bo Lau built a website simulating Elizabeth Holmes’ 2016 Theranos desk, using real documents from the US v. Holmes trial.
-
AMD Unveils MI455X AI Accelerator with 34x Throughput Boost
AMD's MI455X outperforms MI355X by up to 34x in token throughput at high interactivity, according to results from Advancing AI 2026.
-
Anthropic Launches Cyber Mission to Secure Critical Infrastructure
Anthropic announced the Cyber Mission on October 8, 2026, offering tools and resources to secure critical infrastructure and open-source software.
-
Zenity Researchers Hijack All AI Agents in AWS Account
A single prompt allowed Zenity to hijack every AI agent in an AWS account, exposing private data and credentials through vulnerabilities in Amazon's Bedrock AgentCore.
-
Ethereum Researchers Warn AI Could Threaten Wallet Security
Ethereum researchers warn AI math breakthroughs could break wallet security within months, according to a post by Justin Drake.
-
Mathematicians Call for OpenAI Boycott After AI-Generated Proofs Flood Their Field
OpenAI released over 700 AI-generated mathematical papers, sparking calls for a boycott from the Association for Human Mathematics.
-
TechCrunch Disrupt 2026 Offers Last Chance for Discounted Tickets
TechCrunch Disrupt 2026 is set to begin on October 13, with discounted passes available until event doors open.