Tag: model poisoning
5 articles
Popular Python libraries: Stunning Hugging Face danger
Think twice before blindly loading Hugging Face models: researchers found attackers can hide executable Python code in file metadata and malformed pickles so a downloaded model can automatically run malicious payloads. With major libraries and millions of downloads affected, this stealthy supply‑chain trick puts countless projects and machines at risk.
AI Browsers Exclusive: Security Leaders Call Risky
Before you roll out agentic browsers, pause—security leaders warn these AI-powered tools can trade productivity for stealthy new attack surfaces. With embedded models, persistent state and plugins able to act for users, CISOs are being urged to block or tightly control them until hardened safeguards arrive.
PickleScan Exclusive: Critical Flaws Rock AI Supply Chains
Researchers disclosed three critical PickleScan zero-days that let attackers stealthily swap or tamper with local AI models—injecting misinformation, bias, or even exfiltrating data from Python/PyTorch model runners. Exploitable via drive-by browser-origin attacks against assumed-safe local admin endpoints, these flaws show how our trusted AI tooling can become the weakest link in the supply chain.
Gemini AI Exclusive: Dangerous Thinking Robot Malware
What if the AI meant to amplify our thinking could be turned into thinking robot malware that rewrites itself to hide from defenders? New research shows attackers chaining prompt- and log-injection tricks to weaponize Gemini into self-modifying, persistent surveillance agents that sidestep many standard safeguards.
exposed Ollama servers: Risky Must-Have Security Fix
Cisco Talos found 1,100+ publicly exposed Ollama servers, creating easy paths for data theft, malicious model swaps, and other abuse. It’s a wake-up call to fix misconfigurations, enforce authentication, and make secure defaults the norm.