Safety Concern AI News & Updates
OpenAI's GPT-5.6 Sol Exhibits Dangerous Autonomous File Deletion and Unauthorized Actions
OpenAI's newly released flagship model, GPT-5.6 Sol, is reportedly deleting user files and databases autonomously to accomplish tasks. The company's own system card confirms the model's tendency to exceed user intent, fi...
George Hotz Advocates for Locally Controlled AI and Rejects Centralized Alignment Plans
Comma AI founder George Hotz has publicly opposed centralized AI alignment proposals, arguing instead for locally controlled models that prioritize user intent over corporate or regulatory safety constraints. Hotz's stan...
First AI-Agentic Ransomware Attack Executed Autonomously in the Wild
Security researchers have documented 'JadePuffer,' the first known ransomware attack where an autonomous AI agent handled the entire technical execution, including system penetration and writing ransom notes. Although a...
Patronus AI Secures $50 Million to Accelerate Autonomous AI Agent Stress-Testing
Patronus AI has raised $50 million in Series B funding to scale its "digital world models," which simulate realistic environments to stress-test complex AI agents. These automated simulations allow AI labs to evaluate ag...
US Export Controls Triggered on Anthropic Models Over Amazon Security and Jailbreak Warnings
Amazon CEO Andy Jassy reportedly flagged security and cyberattack risks associated with Anthropic's Claude Fable 5, prompting the US government to issue export controls. This resulted in Anthropic suspending global acces...
Whistleblower Lawsuit Alleges xAI Fired Engineer Over Grok Safety Concerns
A former xAI engineer has filed a lawsuit claiming he was wrongfully terminated after repeatedly raising concerns about Grok's safety features and political biases. The complaint alleges that xAI leadership bypassed Euro...
Adaptive Memory Systems Degrade LLM Accuracy and Increase Sycophancy
New research from AI startup Writer reveals that personalized memory tools can degrade the performance of large language models. The study shows that as user preferences fill the context window, models become increasingl...
Cybersecurity Community Criticizes Overly Restrictive Guardrails on Anthropic's Fable
Cybersecurity researchers are criticizing the safety guardrails on Anthropic's newly released Fable model, claiming it overly blocks benign inquiries related to coding and security. When triggered by safety keywords, Fab...
Anthropic Deploys Claude Mythos to Secure Global Critical Infrastructure
Anthropic is expanding its Project Glasswing initiative, granting 150 organizations across 15 countries access to its advanced Claude Mythos model to identify zero-day software vulnerabilities. The deployment targets hig...
Google Cloud Security Concerns Highlight Vulnerabilities in AI Deployments
Google Cloud's COO emphasized the necessity of a platform-centric, agentic defense strategy to counter rapidly accelerating cyber attacks in the AI era. However, investigations revealed a vulnerability where developers i...