LLM security

Articles tagged LLM security on mistr.AI.

• Just 250 Manipulated Documents Are Enough to Make a Large Language Model Vulnerable — Imagine someone being able to sabotage a chatbot with just a few hundred manipulated texts. Anthropic, in collaboration with British security institutes, has found that as few as 250 malicious documents are sufficient to introduce a backdoor into a large language model. The size of the model or the volume of training data makes no difference. What does this mean for AI security?
• OpenAI Launched a Browser with Security Problems It Warns About Itself — OpenAI introduced its new ChatGPT Atlas browser to the public a week ago. Within 72 hours, seven security companies discovered critical vulnerabilities. Atlas fails 94.2% of tests against phishing attacks, while Chrome stops 47%. OpenAI itself warns people not to use Atlas with sensitive data.