Skip to content

Secrets Scanner

This scanner diligently examines user inputs, ensuring that they don't carry any secrets before they are processed by the language model.

Attack scenario

Large Language Models (LLMs), when provided with user inputs containing secrets or sensitive information, might inadvertently generate responses that expose these secrets. This can be a significant security concern as this sensitive data, such as API keys or passwords, could be misused if exposed.

To counteract this risk, we employ the Secrets scanner. It ensures that user prompts are meticulously scanned and any detected secrets are redacted before they are processed by the model.

How it works

While communicating with LLMs, the scanner acts as a protective layer, ensuring that your sensitive data remains confidential.

This scanner leverages the capabilities of the detect-secrets library, a tool engineered by Yelp, to meticulously detect secrets in strings of text.

Types of secrets

  • API Tokens (e.g., AWS, Azure, GitHub, Slack)
  • Private Keys
  • High Entropy Strings (both Base64 and Hex) ... and many more

Usage

from llm_guard.input_scanners import Secrets

scanner = Secrets(redact_mode=Secrets.REDACT_PARTIAL)
sanitized_prompt, is_valid, risk_score = scanner.scan(prompt)

Here's what those options do:

  • detect_secrets_config: This allows for a custom configuration for the detect-secrets library.
  • redact_mode: It defines how the detected secrets will be redacted—options include partial redaction, complete hiding, or replacing with a hash.

Benchmarks

Environment:

  • Platform: Amazon Linux 2
  • Python Version: 3.11.6

Run the following script:

python benchmarks/run.py input Secrets

Results:

Instance Input Length Test Times Latency Variance Latency 90 Percentile Latency 95 Percentile Latency 99 Percentile Average Latency (ms) QPS
AWS m5.xlarge 60 5 2.92 83.84 110.85 132.45 29.75 2016.83
AWS g5.xlarge GPU 60 5 3.34 89.20 118.11 141.23 31.39 1911.67
Azure Standard_D4as_v4 60 5 5.46 114.56 180.92 40.56 421.46 1479.37