Privacy And Redaction For Claude Logs

What To Protect

A Claude log can contain more sensitive information than a normal application log. It may include user prompts, assistant reasoning summaries, file paths, code snippets, terminal output, tool arguments, MCP call details, identity attributes, repository names, customer names, and screenshots or attachments.

Treat raw transcripts as sensitive by default. A transcript that looks harmless at the first prompt can reveal secrets later through a tool result, environment variable, stack trace, config file, or copied support message.

Redact In A Stable Order

First remove secrets and credentials. Then remove personal data, customer identifiers, private repository names, internal hostnames, unreleased product names, and exact file paths if they reveal too much. Finally, review model responses and tool outputs for copied context.

Do not rely only on regexes. They help catch token-shaped strings, but many sensitive facts are ordinary words. A human review step is still required before external sharing.

• Replace secrets with named placeholders such as [REDACTED_API_KEY].
• Keep timestamps and event order unless they reveal private operations.
• Preserve enough structure for a reviewer to understand the incident or bug.

Telemetry Content Controls

OpenTelemetry event export is valuable because it centralizes agent activity. It also expands the blast radius if prompts, responses, and tool inputs are exported without policy. Start with metadata fields and explicitly decide whether content logging is allowed for each environment.

Anthropic documents options for detailed MCP activity and content-related event fields. If those are enabled, route the data to an access-controlled backend with retention and deletion rules, not an open team dashboard.

Safe Sharing Patterns

For a PR comment, share a short derived summary and link to internal evidence. For a vendor support ticket, create a minimal reproduction transcript with secrets removed. For a public blog post, rebuild the scenario with synthetic paths and data.

When in doubt, share the finding, not the raw transcript. The operational value usually comes from sequence, tool category, and final evidence, not from exposing every prompt token.