Promptwall

Key features

• Automatic tools/post-execute inspection for every tool except an explicit trust list, covering canonical values, their independently rendered merge-extensible content blocks, downstream replacements, and additional model contexts.
• Detection is local and pattern-based
• English and Chinese rules for instruction override, role hijack, prompt theft, credential exfiltration, tool coercion, persistence, and obfuscation.
• Quarantine markers that preserve useful surrounding data while removing suspicious instruction spans.
• High-confidence redaction for private keys, AWS/GitHub/Slack/Stripe tokens, JWTs, bearer tokens, and credential assignments.