Newly disclosed vulnerabilities in the Ollama framework allow attackers to execute arbitrary code by feeding the server malicious GGUF model files. This article explains how the mllama metadata parsing flaw works, why versions before 0.7.0 are at risk, and what AI security teams should do to harden their local LLM infrastructure against code execution attacks.
AI-powered fraud is expanding at an unprecedented rate as criminals weaponize generative models, deepfakes and automated identity systems. Organizations must strengthen identity verification and adopt resilience-first strategies to withstand this next-generation threat.
Researchers uncovered serious AI bugs across Meta, Nvidia, Microsoft and open-source inference frameworks after tracking a ShadowMQ deserialization pattern built on ZeroMQ and Python pickle. At the same time, new research shows how Cursor’s AI IDE can be hijacked via rogue MCP servers, turning developer workstations into high-value malware delivery platforms if teams ignore AI supply-chain security.
Researchers hammered ChatGPT, Gemini, and Claude with adversarial prompts across hate, self-harm, sexual content, and crime. Gemini Pro 2.5 leaked the most unsafe answers, Gemini Flash 2.5 resisted best, and Claude models looked strict but cracked under “academic-style” questions in some categories.
Encrypted doesn’t mean invisible. Microsoft’s “Whisper Leak” shows a passive observer can classify AI chat topics by watching packet sizes and timing on streaming language models. Here’s how the side-channel works, who can exploit it, and which mitigations actually move the needle.
OpenAI’s exploration of memory-based ads raises new privacy and compliance risks. Consequently, security teams should treat memory as persistent data, enforce strict prompt hygiene, require explicit consent, and audit retention. Therefore, keep memory off for sensitive workflows until governance, training, and verification land in production.
Agent session smuggling lets a hostile AI agent exploit a live multi-agent conversation, inherit tool authority, and trigger real actions. With scoped credentials, signed steps, and guarded workflows, teams can keep speed without losing control.
Aardvark is OpenAI’s GPT-5–powered security agent that reads repositories like a human, validates exploitability in a sandbox, and proposes precise patches. Consequently, teams can cut detection and remediation time without slowing delivery.
AI-targeted cloaking feeds AI agents a different web than humans see. Learn the risks, detection tactics, and governance steps to keep answers trustworthy.
Agentic AI expands your attack surface because agents read and act on untrusted content. Consequently, indirect prompt injection can hijack tool use, leak data, and trigger risky actions. This guide explains how the attack works, how to detect it, and how to deploy guardrails that actually help at enterprise scale.