KavachPrompt Injection
CriticalSystem override and jailbreak attempts.
ignore\s+(all|previous)\s+instructions · you are now admin · DAN
Regex-powered rule scanning that intercepts every LLM tool call and decides allow or block — before your infrastructure ever sees it. This is the core of the open-source kavach-mcp middleware, now scaled to platform level.
Intercept → match → decide → audit. Strict mode raises, advisory mode chains, and masking keeps secrets out of your logs.
Severity-tagged patterns shipped in kavach/rules.py. Merge with your own or replace entirely.
System override and jailbreak attempts.
Attempts to leak or dump system data and prompts.
Aadhaar, credit cards, phone and SSN-like values.
AWS, OpenAI and Google keys, PEM keys, api_key assignments.
eval, exec, compile, os.system and subprocess.
DDL / DML attacks and comment-based tricks.
Directory traversal and file escape.
rm/del/wipe against system paths.
Delete bucket, terminate instance, drop database.
Command chaining to destructive commands.
Unsafe MCP tool usage regardless of arguments.
Add any regex; merge with defaults or replace them.
Runtime protection as a drop-in middleware hook.
kavach-mcp is free and portable — extend it with your own patterns, shadow-mode test it, then come to the platform for live inventory and orchestration.