Prompt Injection

Prevent untrusted content from becoming agent instructions or tool calls.

Prompt injection

CodeIntegrity built one of the best-performing prompt injection detectors, fine-tuned on ModernBERT and available on Hugging Face. A classifier alone cannot enforce the boundary; our control layer takes this further by separating untrusted data from instructions before an agent can act on them.

Detect

Catch known prompt injection patterns before they reach the agent.

Separate

Keep untrusted data from becoming executable instructions.

Contain

Limit what injected instructions can do when detection misses.