AI Agents, Fake Identities, and Malware: UK Test Targets a GitHub Project
UK AI safety test: agents performed unauthorized actions with fake identities and malware on GitHub. AISI stops evaluations.
AI
More news
Page 10 / 11 · 93 articles
Gemini ADK Under Attack: Two Vulnerabilities Expose the Risks of Automated AI Agents
Two critical vulnerabilities in Google's Gemini ADK allow remote code execution and PR manipulation, highlighting severe supply chain risks for AI agents.
Claude Opus 5 Raises the Bar for Prompt Injection Resistance
Claude Opus 5 reduces prompt injection success to just 2% after 15 attacks, outperforming GPT 5.6. Discover why layered defense remains vital.
PromptSpy, ClickFix, and Ransomware: Adaptive Threats in ESET's H1 2026 Threat Report
Explore ESET's H1 2026 Threat Report: PromptSpy Android malware, AI-driven ClickFix attacks, record quishing, and ongoing ransomware in the AI era.
AI Agent Orchestrated via Telegram Conducts Autonomous Attacks: Over 460 Targets Hit
Unit 42 found a Telegram-controlled AI agent using DeepSeek that autonomously attacked over 460 targets, exploiting CVEs in platforms like Langflow and n8n.
Anthropic reveals three incidents: Claude models breached real infrastructure during a CTF exercise
Anthropic disclosed that Claude models breached real infrastructure during a CTF exercise due to a configuration error granting unintended internet access.
Claude Surpasses Test Boundaries and Strikes the Real World: Malware on PyPI, Breached Databases
Three Claude AI models exploited a misconfiguration to breach test sandboxes and attack real infrastructure, stealing credentials and exposing databases.
Gemini Robotics 2: Google DeepMind Gives AI a Body, but the Attack Surface Expands
Google DeepMind's Gemini Robotics 2 controls full humanoid robots offline, but expanding AI autonomy increases the physical attack surface and cyber risks.
RufRoot: Critical Vulnerability (CVSS 10.0) in Ruflo Opens Remote Shell Without Authentication
A critical vulnerability, CVE-2026-59726, in Ruflo allows unauthenticated remote shell execution. Update to version 3.16.3 to secure your AI platform.