Hidden Instructions Detected

Content contains patterns that may manipulate agent behavior

Quarantined

What Was Found

The scanner detected patterns commonly used in prompt injection attacks or attempts to override an agent's instructions. This could include hidden system prompts, jailbreak phrases, or role manipulation attempts.

Why This Matters

Prompt injection is one of the most significant risks to AI agents. Malicious instructions hidden in memory could cause your agent to:

Ignore their safety guidelines or boundaries
Reveal sensitive information
Take actions against their configured purpose
Adopt a different persona without your knowledge

This content has been quarantined for your review. You can approve it if you're certain it's safe, or reject it to keep your agent protected.

Common Causes (May Be False Positive)

Conversations about AI safety or jailbreaking (discussing the topic, not attempting it)
Creative writing that includes "system prompt" style formatting
Technical documentation about prompt engineering
Roleplay scenarios with persona instructions
Invisible Unicode characters (sometimes added by text editors)

What You Can Do

Review the content in the Quarantine section:

Approve if you recognize the content and trust its source
Reject if you don't recognize it or it seems suspicious
When in doubt, reject — you can always re-import trusted content later

If this is from a conversation about prompt injection (rather than an actual attack), it's likely safe to approve. Context matters.

Review Quarantine Back to Dashboard