Model Self-Correction Vulnerabilities

The discovery that 'system prompts'—the hidden instructions provided by developers—can be extracted or overridden by advanced prompt injection attacks.

Sources

Open the full topic