AI researchers have discovered numerous ways to bypass the safety rules of Bard and ChatGPT, indicating virtually unlimited possibilities.

AI researchers have discovered numerous ways to bypass the safety rules of Bard and ChatGPT, indicating virtually unlimited possibilities.

Researchers Find Ways to Break Through Guardrails on AI-powered Chatbots

Researchers at Carnegie Mellon University and the Center for A.I. Safety have discovered potentially unlimited ways to bypass safety guardrails on major AI-powered chatbots. These guardrails, implemented by tech companies like OpenAI, Google, and Anthropic, are designed to prevent the chatbots from generating harmful or malicious content.

Large language models such as ChatGPT, Bard, and Anthropic’s Claude rely heavily on moderation to ensure they cannot be used for nefarious purposes. These models are equipped with comprehensive safety measures to prevent them from providing instructions on building bombs or engaging in hate speech.

However, a recent report released by the researchers demonstrated how jailbreaks developed for open-source systems could be used to target mainstream and closed AI systems. By using automated adversarial attacks, such as adding characters to the end of user queries, the guardrails could be overridden, leading the chatbots to produce harmful content, misinformation, or hate speech.

What makes the researchers’ approach unique is the fully automated nature of their hacks, allowing for the potential creation of a “virtually unlimited” number of similar attacks. Unlike previous jailbreaks, their method does not rely on manual intervention and can be executed automatically.

The researchers promptly disclosed their findings to Google, Anthropic, and OpenAI. A Google spokesperson stated that while this is an issue affecting all large language models (LLMs), important guardrails have been implemented in Bard, and further improvements will be made over time.

• Harvard leadership expert says Elon Musk is ‘out of his eleme...
• July 2023’s Best High-Yield Savings Rates (Up to 5.25% Earned)
• July 2023’s Best CD Rates (Up to 7.19% APY)

OpenAI and Anthropic have yet to respond to the researchers’ disclosure, as Insider’s request for comment was made outside of normal working hours.

When OpenAI’s ChatGPT and Microsoft’s AI-powered Bing were initially released, users took delight in finding ways to undermine the systems’ guidelines. Early hacks, including one that tricked the chatbot into answering without content moderation, were swiftly patched up by the tech companies.

However, the researchers pointed out that it remains “unclear” whether such behavior can ever be fully blocked by the companies behind the leading models. This raises questions about the effectiveness of AI system moderation and the safety implications of releasing powerful open-source language models to the public.

In conclusion, researchers have discovered potential vulnerabilities in the guardrails implemented by major AI-powered chatbots. This brings into focus the need for continuous improvement in safety measures and raises concerns about the effectiveness of moderation protocols. As the field of AI continues to evolve, it will be crucial to address these challenges to ensure the responsible and secure use of AI-powered technologies.

• ai
• Bing
• Chatbots