Secret Keys

Identifies potential API keys, secrets, and credentials in text using entropy analysis and pattern matching. Scans text for high-entropy strings that look like secrets, uses pattern matching to identify common secret formats, applies entropy analysis to detect random-looking strings, and helps prevent accidental exposure of sensitive credentials.

Supported stages

Stage	Supported	Notes
pre_flight	✅	Catch secrets accidentally pasted into user messages
input	✅	Runs concurrently with the LLM call
output	✅	Recommended. Prevents the LLM from leaking secrets it was given in the system prompt or context

Configuration

{
    "name": "Secret Keys",
    "config": {
        "threshold": "balanced",
        "custom_regex": ["my-custom-[a-zA-Z0-9]{32}", "internal-[a-zA-Z0-9]{16}-key"]
    }
}

Parameters

• threshold (optional): Detection sensitivity level (default: "balanced")
  • "strict" - Most sensitive, may have more false positives (commonly flag high entropy filenames or code)
  • "balanced" - Default setting, balanced between sensitivity and specificity
  • "permissive" - Least sensitive, may have more false negatives
• custom_regex (optional): List of custom regex patterns to check for secrets

Mend Platform configuration

When a policy is loaded from the Mend Platform API the secrets guardrail is identified by "id": "secrets". The API currently carries no per-row config object for this guardrail, so it is always loaded with the SDK defaults (threshold="balanced", custom_regex=null).

Example API response row:

{
    "id": "secrets",
    "action": "Alert",
    "status": true,
    "direction": "Input",
    "guardrail": "Secrets"
}

To use a non-default threshold or custom patterns, configure the guardrail through an explicit SDK policy (see Configuration above) rather than relying on the platform policy row.

Supported secrets

Detection uses two complementary approaches applied to every whitespace-separated token in the text.

Known-prefix matching

Any token whose prefix matches a well-known credential format is flagged immediately, regardless of entropy. The recognised prefixes are:

Prefix	Credential type
sk-	OpenAI secret keys
sk_	Stripe secret keys
pk_	Stripe publishable keys
pk-	Generic private/public API keys
ghp_	GitHub Personal Access Tokens
AKIA	AWS Access Key IDs
xox	Slack tokens (xoxb-, xoxp-, xoxa-, …)
SG.	SendGrid API keys
hf_	Hugging Face access tokens
key-	Generic API keys
api-	Generic API keys
apikey-	Generic API keys
token-	Generic tokens
secret-	Generic secrets
SHA:	SHA digests used as secrets
Bearer	HTTP Bearer auth tokens

Entropy-based heuristics

Tokens that do not match any known prefix are still flagged if they meet all of the threshold criteria simultaneously:

Criterion	strict	balanced (default)	permissive
Minimum length (chars)	10	15	20
Minimum Shannon entropy (bits)	3.5	3.8	4.0
Minimum character-type diversity¹	2	3	3
Ignore file extensions / URLs	No	Yes	Yes

¹ Character-type diversity counts how many of the four character classes are present in the token: lowercase, uppercase, digits, and special characters. A value of 3 means the token must use at least three of these classes.

False-positive suppression

In balanced and permissive modes the guardrail skips tokens that look like file paths (e.g. report.pdf, config.yaml) or HTTP/HTTPS URLs, since those rarely contain actual secrets.

Custom patterns

Use custom_regex to flag credentials that follow a project-specific format. Custom patterns are checked first and bypass the entropy heuristics:

{
    "name": "Secret Keys",
    "config": {
        "custom_regex": ["int-[a-zA-Z0-9]{32}", "corp_[A-Z0-9]{16}"]
    }
}

Implementation Notes

• Pre-configured Sensitivity: Threshold values automatically set appropriate entropy, length, and diversity requirements
• Pattern Matching: Looks for common secret prefixes and formats

What It Returns

Returns a GuardrailResult with the following info dictionary:

{
    "guardrail_name": "Secret Keys",
    "detected_secrets": ["sk-abc123...", "Bearer xyz789..."]
}

• detected_secrets: List of potential secrets detected in the text