OpenAI — ChatGPT Lockdown Mode
AI relevance: Lockdown Mode constrains networked agent features in ChatGPT to reduce prompt-injection driven data exfiltration risks.
- OpenAI introduced an optional Lockdown Mode for high-risk users to deterministically disable risky capabilities.
- In Lockdown Mode, web browsing is limited to cached content so no live outbound requests leave OpenAI’s network.
- High-risk tools like Agent Mode, Deep Research, file downloads, and Canvas networking are disabled.
- Image responses are blocked, but users can still upload images and use image generation.
- Apps/connectors aren’t disabled by default; admins must explicitly scope app actions to avoid exfil paths.
- OpenAI added “Elevated Risk” labels to features with known security tradeoffs across ChatGPT, Atlas, and Codex.
- Availability starts with Enterprise, Edu, Healthcare, and Teachers, with consumer rollout planned.
Why it matters
- Prompt injection is an operational risk once LLMs can browse, call tools, or access internal systems.
- Deterministic controls matter for high-risk roles where data exfiltration is the main threat.
- Enterprise AI governance needs feature-level guardrails, not just model-level policies.
What to do
- Enable Lockdown Mode for high-risk users (execs, security teams, sensitive workflows).
- Audit app/connectors and restrict write actions to trusted destinations only.
- Document which features are Elevated Risk and align them with internal data policies.
- Test prompt-injection scenarios with Lockdown on/off to measure residual risk.