OpenAI Unveils Lockdown Mode to Protect Sensitive Data from Prompt Injection (techcrunch.com)

OpenAI has introduced a new feature called Lockdown Mode aimed at enhancing protection against prompt injection attacks, where adversarial instructions are concealed in online content. This mode disables live web browsing and the retrieval of images from the web, allowing access only to cached content, while also limiting deep research and agent modes. Although Lockdown Mode provides an extra layer of security, OpenAI warns that vulnerabilities may still exist, particularly with cached web data and uploaded files potentially affecting response accuracy. The significance of Lockdown Mode lies in its targeted approach to safeguarding sensitive data for businesses and organizations that are particularly at risk of data leaks. By implementing this feature, OpenAI acknowledges the ongoing challenges of prompt injection vulnerabilities while striving to mitigate the risk of unintentional data exfiltration. Currently, Lockdown Mode is being rolled out to self-serve ChatGPT Business accounts and selected personal accounts, catering specifically to those needing stricter data protection measures.