Anthropic has introduced a new AI model it considers too dangerous to release publicly. The model, called Claude Mythos Preview, can autonomously find and exploit security vulnerabilities in software. Instead of making it widely available, Anthropic is sharing access with a coalition of more than 40 organizations as part of an initiative called Project Glasswing. …
OpenAI has removed the word “safely” from its mission statement, a change documented in its 2024 tax filing with the Internal Revenue Service. The deletion coincides with the company’s transformation from a nonprofit organization into a for-profit business. The original mission statement from 2022 and 2023 read: “to build general-purpose artificial intelligence (AI) that safely …
Anthropic has emerged as a formidable competitor in the artificial intelligence industry by focusing on enterprise customers and positioning itself as the most safety-conscious AI company. The approach appears to be paying off both commercially and in investor confidence, even as critics question whether the company can maintain its principles while racing to capture market …
Two major security incidents have exposed vulnerabilities in the rapidly growing ecosystem around AI agents, revealing risks when artificial intelligence creates software without human oversight. Cybersecurity firm Wiz discovered a significant security flaw in Moltbook, a social network designed exclusively for AI agents, Raphael Satter reports for Reuters. The vulnerability exposed private messages between agents, …
Security researchers have identified critical vulnerabilities in OpenClaw, an open-source AI assistant that automates email, calendar management, and other tasks. The project rebranded from Clawdbot to Moltbot and then to OpenClaw after receiving a trademark complaint from Anthropic. The core problem lies in Model Context Protocol, the framework OpenClaw uses to connect with various services. …
Microsoft is encountering significant pushback from security experts and users regarding its strategy of integrating advanced artificial intelligence into its Windows operating system. The criticism centers on a new experimental feature called Copilot Actions, which Microsoft itself has warned could expose users to malware and data theft. The company introduced Copilot Actions as a set …
AI-powered browsers designed to automate web tasks can be hijacked through hidden instructions embedded in websites, creating a significant security risk. Harshith Vaddiparthy reports for VentureBeat that these tools can be tricked into executing harmful commands without the user’s knowledge. The AI browser Comet from Perplexity serves as an example of this vulnerability. The core …
AI-powered chatbots often provide incorrect website addresses for major companies, creating a new attack vector for criminals. According to a report by threat intelligence firm Netcraft, this vulnerability can be exploited for sophisticated phishing schemes. The findings were detailed in an article by Iain Thomson for The Register. Netcraft researchers tested GPT-4 models by asking …
AI safety researchers have created the first benchmark specifically designed to detect manipulative behaviors in large language models, following a concerning incident with ChatGPT-4o’s excessive flattery toward users. Leon Yen reported on the development for VentureBeat. The DarkBench framework, developed by Apart Research founder Esben Kran and collaborators, identifies six categories of problematic AI behaviors. …
OpenAI has published a detailed explanation about the technical issues that caused GPT-4o to become overly sycophantic in April. In a comprehensive blog post, the company revealed that an update rolled out on April 25 made the model excessively eager to please users by validating doubts, fueling anger, and reinforcing negative emotions in unintended ways. …