We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
OpenAI Blog·Feb 13·5 min read
Introducing Lockdown Mode and Elevated Risk labels in ChatGPT
TL;DR
Organizations using ChatGPT face prompt injection attacks (malicious inputs that manipulate AI behavior) and data exfiltration risks. OpenAI introduced Lockdown Mode, a restrictive operational state, and Elevated Risk labels to flag suspicious interactions.
✦ Why It Matters
Engineers can now implement layered defenses against prompt injection and data theft in production ChatGPT deployments.
Key Takeaways
- Enable Lockdown Mode for workflows handling sensitive data or high-risk use cases.
- Monitor Elevated Risk labels in audit logs to detect and respond to suspicious interactions.
- Combine Lockdown Mode restrictions with label-based alerting for defense-in-depth security.
- Test Lockdown Mode impact on legitimate workflows before production deployment.
Full Summary
ChatGPT deployments in enterprise environments are vulnerable to prompt injection—techniques where attackers craft inputs to override intended system behavior and extract sensitive data. OpenAI released two complementary security features: Lockdown Mode, which restricts ChatGPT's capabilities and external integrations to minimize exposure to malicious prompts, and Elevated Risk labels, automated flags that surface potentially dangerous interactions to security teams.
Lockdown Mode operates as a hardened configuration reducing feature surface area, while labels leverage pattern detection to identify suspicious activity. Organizations can enable Lockdown Mode for high-sensitivity workflows and monitor labeled interactions through audit logs.
These tools shift security responsibility toward detection and containment rather than prevention alone, allowing teams to balance usability with risk mitigation.