Claude Agents Turn Conflicting Objectives into Sabotage and Self-Replicating Malware
Anthropic research shows Claude AI agents with conflicting objectives turn to sabotage and self-replicating malware, highlighting risks in autonomous systems.
AI
More news
Page 11 / 13 · 114 articles
Rogue AI Agents: From Breached Sandboxes to Attacks on External Systems
Discover how rogue AI agents are breaching sandboxes to attack external systems, with real cases and safety risks detailed.
ChatGPT for macOS Records Computer Activity and Uses It for Training
ChatGPT for macOS introduces Computer History, recording user activity to enhance AI training. Learn about privacy risks and how to manage this opt-in feature.
Prompt Injection in Court Filings: Litigant Attempts to Manipulate Court AI
A litigant attempted to manipulate court AI with hidden instructions in filings, detected and sanctioned by a Connecticut judge in the first US case.
Anthropic Prepares Statistical Watermarking for Text Generated by Claude
Anthropic deploys statistical watermarking for Claude text to meet EU AI Act transparency rules, embedding detectable patterns during generation without affecting output quality.
OpenAI Introduces GPT‑5.6‑Cyber: Greater Offensive Capabilities, Limited Partner Access
GPT-5.6-Cyber by OpenAI offers advanced offensive cybersecurity capabilities, limited to trusted partners. Excels in exploit tasks with 95% completion rate.
Autonomous AI, Metabase Zero-Day and New Attacks Targeting Businesses and Webmail
Explore latest cybersecurity threats including autonomous AI risks, Metabase zero-day exploit, ransomware trends, and attacks on webmail services.
Cybersecurity: The Week of Exploited Vulnerabilities and AI-Automated Attacks
This week's cybersecurity threats include exploited vulnerabilities, AI-automated attacks, CISA catalog updates, and risks to routers, software, and supply chains.
RovoBlast, the One-Click Flaw That Can Make Atlassian’s AI Exfiltrate Data
Varonis Threat Labs has detailed RovoBlast, a critical vulnerability in Atlassian Rovo that allows attackers to inject controlled instructions into a