Objective

Implement optional DLP (Data Loss Prevention) scanning in Squid to detect and prevent exfiltration of API keys, tokens, and credentials in outgoing requests.

Context

Current state: Domain allowlisting restricts which hosts can be contacted, but doesn't inspect request content.

Risk: Attacker could encode sensitive data (API keys, tokens) in HTTP requests to allowed domains (e.g., creating GitHub gists with credentials).

Risk level: 🟡 MEDIUM - Information disclosure via allowed domains

Implementation Approach

1. Add --enable-dlp flag to enable content inspection (opt-in for performance)
2. Define regex patterns for common credential formats:
   • GitHub tokens: ghp_[A-Za-z0-9]{36}, gho_[A-Za-z0-9]{36}, ghs_[A-Za-z0-9]{36}
   • OpenAI API keys: sk-[A-Za-z0-9]{48}
   • AWS keys: AKIA[0-9A-Z]{16}
   • Generic patterns: [Aa]pi[_-]?[Kk]ey, [Tt]oken
3. Use Squid's adaptation_service_set with ICAP or eCAP adapter
4. Log blocked requests with [DLP_BLOCKED] prefix
5. Return 403 Forbidden when sensitive pattern detected

Files to Modify

• src/cli.ts - Add --enable-dlp flag
• src/squid-config.ts - Generate ICAP/eCAP configuration when enabled
• src/types.ts - Add DLP config to WrapperConfig
• containers/squid/dlp-adapter.sh - Simple ICAP adapter script for pattern matching
• containers/squid/Dockerfile - Install ICAP adapter dependencies
• README.md - Document DLP feature and detected patterns

Testing

• Test detection of GitHub personal access token (ghp_...)
• Test detection of OpenAI API key (sk-...)
• Test detection of AWS access key (AKIA...)
• Verify legitimate requests without credentials pass through
• Confirm DLP only active when --enable-dlp flag used

Acceptance Criteria

• --enable-dlp flag enables content inspection
• GitHub tokens detected and blocked
• OpenAI API keys detected and blocked
• AWS credentials detected and blocked
• Blocked requests logged with [DLP_BLOCKED] prefix
• Performance impact documented (< 10% latency increase)
• Feature documented with security benefits
  Related to [plan] security improvements from threat model review #306

AI generated by Plan Command for discussion #302