AI security

Noma Labs discovered a critical flaw in Docker's Ask Gordon AI assistant allowing attackers to hijack AI reasoning through malicious image metadata, leading to remote code execution or data exfiltration.

CVE-2026-22778, a critical RCE in vLLM versions 0.8.3-0.14.0, chains a PIL information leak with a JPEG2000 heap overflow to achieve code execution through a malicious video link.

CVE-2026-25253 (CVSS 8.8) allows attackers to steal authentication tokens and achieve RCE through a single malicious link via cross-site WebSocket hijacking—even on localhost-only OpenClaw instances.

Over 400 malicious OpenClaw AI agent skills on ClawHub deploy Atomic Stealer via ClickFix-style social engineering. The hightower6eu account alone published 314 malicious skills targeting crypto and developer credentials.

Ranking the best email security platforms based on the 2025 Forrester Wave results, AI-powered detection, and defense against phishing, BEC, and multi-channel social engineering.

Ranking the leading network detection and response platforms based on AI-driven threat detection, encrypted traffic analysis, cloud network visibility, and SOC integration.

A practical guide to securing AI and large language model applications, covering the OWASP Top 10 for LLMs (2025), prompt injection defenses, RAG security, AI agent risks, and compliance with NIST AI RMF and the EU AI Act.

A practical guide to securing large language model and AI deployments, covering prompt injection, data extraction, RAG pipeline security, AI gateways, input/output filtering, and the OWASP Top 10 for LLM Applications.

Practical guide to implementing NIST's AI Risk Management Framework (AI RMF 1.0) and the Cyber AI Profile (IR 8596), covering the Govern, Map, Measure, and Manage functions for organizations building or deploying AI systems.

Microsoft introduces new Purview DLP integration, sensitivity label enforcement, and oversharing assessment tools for Copilot for Microsoft 365, responding to widespread CISO concerns about AI assistants accessing sensitive data through existing permissions.

Varonis discovered a prompt injection attack chain that could steal sensitive data from Microsoft Copilot with a single click, bypassing safety filters through double-request and chain-request techniques. Patched January 13, 2026.

Autonomous cybersecurity platform delivering AI-powered endpoint, cloud, and identity protection with automated response capabilities.