Regex Prompt Guard Policy
Policy Name |
Regex Prompt Guard |
Summary |
Blocks LLM requests that match deny-list regex patterns |
Category |
LLM |
First Omni Gateway version available |
v1.11.4 |
Release Notes |
|
Returned Status Codes |
403 - Forbidden: Request blocked because prompt content matched one or more denied patterns |
Summary
| This policy supports OpenAI and Anthropic input formats. |
The Regex Prompt Guard policy scans LLM prompts and blocks requests when prompt content matches regex deny-list patterns.
The policy uses the Rust regex crate which supports RE2-style expressions to guarantee linear-time matching and eliminate the backtracking risk.
For example, you can use the policy to block attacks such as:
-
Prompt injection
-
Jailbreak attempts
-
SQL injection patterns
-
HTML script injection
Configuring Policy Parameters
Managed Omni Gateway and Omni Gateway Connected Mode
When you apply the policy from the UI, the following parameters are displayed:
| Element | Required | Description |
|---|---|---|
Pattern Name |
Yes |
Human-readable pattern name. The policy includes this value in the blocked-response details. |
Regex Pattern |
Yes |
RE2-style regex expression used to match prompt content. Use |
Response on Block
When a request is blocked, the policy returns a response body containing the matched deny patterns:
{
"error": "Request blocked - content matches denied pattern",
"matched_patterns": [
"prompt-injection"
]
}Regex Support
The Regex Prompt Guard Policy uses the Rust regex crate which supports RE2-style expressions to guarantee linear-time matching and eliminate the backtracking risk. This ensures that customer-supplied regex patterns can’t negatively impact gateway performance or availability.
The Regex Prompt Guard policy doesn’t support Perl Compatible Regular Expressions (PCRE), but supports these PCRE features:
-
Standard character classes
-
Alternation
-
Grouping
-
Greedy or lazy quantifiers
-
Anchors
-
Inline flags, such as
(?i), for case-insensitivity.
These PCRE-specific features aren’t supported:
-
Lookaround assertions:
(?=…),(?!…),(?⇐…),(?<!…) -
Backreferences:
\1,\2,(?P=name) -
Recursive or subroutine calls:
(?R),(?1) -
Possessive quantifiers:
a*`, `a+ -
Special constructs:
\G,\K, and conditional patterns(?(1)yes|no)
See Also
-
Creating and Managing Model Proxies - Overview of model proxies and routing
-
LLM Token Based Rate Limit Policy - Token-based rate limiting for model proxies