OpenAI is launching two new security features in ChatGPT to address growing threats to its AI systems, according to a recent blog post.
As AI services increasingly connect to wider parts of the web and more external apps, the risk of so-called “prompt injection attacks” also increases. A prompt injection attack is when someone crafts a deceptive prompt in an attempt to trick the LLM into following malicious instructions and/or revealing sensitive information.
One of the new features in ChatGPT is Lockdown Mode, an optional security mode aimed at users with high privacy requirements. This mode strictly limits how ChatGPT interacts with external systems. Certain tools and features are completely disabled, and web browsing is only allowed via cached content instead of direct network calls. Lockdown Mode will first be available to enterprise customers and will later be launched to consumers in the coming months.
At the same time, clearer risk labeling will be introduced, with a uniform label bearing the text “Elevated Risk” for features that pose an increased security risk (for example, those that give AI tools network access). The labels will be visible in ChatGPT, ChatGPT Atlas, and Codex.