Tag: AI Security
Claude models attacked real companies they mistook for a game
Anthropic found three incidents where its Claude models reached live production systems during capture-the-flag evaluations.
CVSS 10 Ruflo bug gives attackers full control of AI agents and their memory
Ruflo's unauthenticated MCP bridge earns a perfect CVSS 10 as researchers demonstrate AI memory poisoning.
Microsoft launches first cybersecurity AI model as part of Project Perception
Microsoft unveiled MAI-Cyber-1-Flash, its first specialized cybersecurity AI model, scoring 96% on CyberGym benchmarks and cutting vulnerability discovery costs by…
Google restricts powerful new Gemini Cyber AI to government partners only
Google DeepMind released Gemini 3.5 Flash Cyber, a specialized AI for vulnerability hunting, restricted to governments and trusted partners due…
OpenAI models break out of sandbox and autonomously hack Hugging Face servers
OpenAI confirmed its GPT-5.6 Sol and another pre-release model escaped the sandbox and autonomously hacked Hugging Face's servers.
Claude for Chrome flaw lets rogue extensions read Gmail messages
Researchers discovered that malicious Chrome extensions can abuse a permission flaw in the Claude for Chrome extension to read Gmail…
Microsoft patches record 570 flaws as AI accelerates vulnerability discovery
Artificial intelligence helped Microsoft discover and patch 570 security vulnerabilities in July, nearly tripling last month's record-breaking patch count.
Memghost attack plants false memories in AI agents through a single email
Researchers discovered the MemGhost technique, which injects persistent false memories into AI agents via crafted email content, causing long-term manipulation…