Skip to content

masking.md

Latest commit

 

History

History
121 lines (80 loc) · 3.51 KB

masking.md

File metadata and controls

121 lines (80 loc) · 3.51 KB
Raw

Data Masking

English | 简体中文

LoongSuite Pilot can mask common secrets before normalized events are sent to output backends. Use this when prompts, completions, tool arguments, or tool results may contain credentials.

Masking is separate from message content capture:

  • captureMessageContent: false reduces whether full message or tool content is collected.
  • mask scans configured output fields and replaces high-confidence secrets that are still present.

For sensitive environments, use both.

How To Enable Masking

Config file:

{
  "mask": {
    "mode": "all"
  }
}

Or, Environment variable:

export LOONGSUITE_PILOT_MASK_MODE=all

Restart Pilot after changing config:

loongsuite-pilot restart

Mask Modes

Mode Behavior
none Do not mask fields. This is the default when no mask mode is configured.
all Enable all built-in sensitive data rules.
custom Enable only the mask types listed in mask.types.

Mask Types

Type What It Covers
cloudAccessKey Alibaba Cloud, AWS, and Tencent-style access key IDs.
apiKey OpenAI-compatible and GitHub-style API keys.
privateKey PEM or OpenSSH private key blocks.
databaseUrl Database URLs with embedded passwords.

Custom mode example:

{
  "mask": {
    "mode": "custom",
    "types": ["apiKey", "databaseUrl"]
  }
}

Equivalent environment variables:

export LOONGSUITE_PILOT_MASK_MODE=custom
export LOONGSUITE_PILOT_MASK_TYPES=apiKey,databaseUrl

Replacement Values

When a rule matches, Pilot replaces the secret with a fixed marker:

Mask Type Replacement Marker
cloudAccessKey [ACCESSKEY_MASKED]
apiKey [APIKEY_MASKED]
privateKey [PRIVATEKEY_MASKED]
databaseUrl [DATABASEURL_MASKED]

The original value is not preserved in the emitted event.

What Gets Masked

Masking runs in the collector before events are written to JSONL, SLS, HTTP, or OTLP trace output.

Pilot focuses masking on fields that may contain user or tool content, such as:

  • LLM input and output messages.
  • Tool call arguments.
  • Tool call results.
  • Known agent-specific content fields, such as agent.content, agent.inline_diff_message, and selected compact content fields.

Stable metadata such as model names, token counts, durations, Git branch, and workspace path is not intended to be scanned as secret-bearing content.

Verify Masking

  1. Enable JSONL output and masking.
  2. Restart Pilot.
  3. Trigger an agent event that contains a known test secret pattern.
  4. Inspect local output:
tail -f ~/.loongsuite-pilot/logs/output/*.jsonl

Expected output should contain mask markers such as [APIKEY_MASKED] instead of the original secret.

Related Configuration