AI Safety Guardrails Crumble Under Advanced Attacks

A new study has revealed that the safety guardrails on advanced AI models are failing to prevent attacks, raising alarms in the crypto and tech sectors. Researchers at Nature found that four large reasoning models were able to bypass security filters with a remarkable 97.14% success rate against nine frontier AI models from companies like OpenAI, Anthropic, and Google.

The research used simple prompts and strategic persuasion techniques in multi-turn conversations to test the limits of these models. DeepSeek-R1 was particularly successful, achieving a 100% attack success rate on HarmBench prompts in Cisco-linked testing.

Additionally, security firm HiddenLayer has documented universal bypass techniques that work across multiple leading large language models, including GPT-4 and Claude. This raises concerns about the commercialization of AI attacks, with a Russian-speaking threat actor operating under the name 'Trim' packaging jailbreak techniques into a paid offensive security platform.

The study's findings highlight the need for improved security measures to protect against these advanced threats. As AI technology continues to advance, it is clear that more robust safety protocols are required to prevent attacks and maintain trust in these systems.