Skip to content

security-and-privacy.md

Latest commit

 

History

History
52 lines (31 loc) · 4.63 KB

security-and-privacy.md

File metadata and controls

52 lines (31 loc) · 4.63 KB

Security & Privacy

Trace stores AI session transcripts and metadata in your git repository. This document explains what data is stored, how sensitive content is protected, and how to configure additional safeguards.

Transcript Storage & Git History

Where data is stored

When you use Trace with an AI agent (Claude Code, Codex, Gemini CLI, OpenCode, Cursor, Factory AI Droid, Copilot CLI), session transcripts, user prompts, and checkpoint metadata are committed to a dedicated branch in your git repository (trace/checkpoints/v1). This branch is separate from your working branches, your code commits stay clean, but it lives in the same repository.

Trace also creates temporary local branches (e.g., trace/<short-hash>) as working storage during a session. These shadow branches store file snapshots and transcripts without redaction. They are cleaned up when session data is condensed (with redaction) into trace/checkpoints/v1 at commit time. Shadow branches are not pushed by Trace — do not push them manually, as unredacted content would be visible on the remote.

Anyone with access to your repository can view the transcript data on the trace/checkpoints/v1 branch. This includes the full prompt/response history and session metadata. Note that transcripts capture all tool interactions — including file contents, MCP server calls, and other data exchanged during the session.

If your repository is public, this data is visible to the trace internet.

What Trace redacts automatically

Trace automatically scans transcript and metadata content before writing it to the trace/checkpoints/v1 branch. Five secret detection methods run during condensation:

  1. Entropy scoring — Identifies high-entropy strings (Shannon entropy > 4.5) that look like randomly generated secrets, even if they don't match a known pattern.
  2. Pattern matching — Uses Betterleaks built-in rules to detect known secret formats.
  3. Credentialed URI detection — Redacts URLs with embedded passwords, such as scheme://user:password@host.
  4. Database connection-string detection — Redacts JDBC, Postgres keyword DSN, SQL Server, and ODBC-style connection strings containing passwords.
  5. Bounded credential value detection — Redacts password-like config values such as DB_PASSWORD=... and PGPASSWORD=... while preserving the surrounding key.

Detected secrets are replaced with REDACTED before the data is ever written to a git object. This is always on and cannot be disabled.

Recommendations

If your AI sessions will touch sensitive data:

  • Use a private repository. This is the simplest and most complete protection. Transcripts on trace/checkpoints/v1 are only visible to collaborators.
  • Avoid passing sensitive files to your agent. Content that never enters the agent conversation never appears in transcripts.
  • Review before pushing. You can inspect the trace/checkpoints/v1 branch locally before pushing it to a remote.

What Gets Redacted

Secrets (always on)

Betterleaks pattern matching covers cloud providers (AWS, GCP, Azure), version control platforms (GitHub, GitLab, Bitbucket), payment processors (Stripe, Square), communication tools (Slack, Discord, Twilio), private key blocks (RSA, DSA, EC, PGP), and generic credentials (bearer tokens, basic auth, JWTs). Dedicated credentialed URI detection covers URLs that embed passwords. Additional database connection-string detection covers DB DSNs and query-parameter passwords not reliably covered by generic secret rules. Entropy scoring catches secrets that don't match any known pattern.

All detected secrets are replaced with REDACTED.

To reduce over-redaction, Trace preserves structural transcript fields such as IDs and paths, ignores common placeholder values, and redacts only credential values for bounded key/value forms. When a connection string contains a real (non-placeholder) password, it is redacted as a unit because partial fragments can still expose sensitive material; connection strings whose passwords are placeholders (e.g. ${DB_PASSWORD}) are left intact.

Limitations

  • Best-effort. Novel or low-entropy secrets (short passwords, predictable tokens) may not be caught.
  • Filenames and binary data. Secrets in filenames, binary files, or deeply nested structures may not be detected.
  • JSONL skip rules. Trace skips scanning fields named signature or ending in id/ids, and objects whose type starts with image or equals base64, to avoid false positives.
  • Users are ultimately responsible for reviewing what they commit and push. Redaction is a safety net, not a guarantee.