September 2026 Summaries

AI agent traces provide valuable debugging context by recording model inputs and outputs, tool calls, arguments, results, errors, and application metadata, but this visibility can also expose credentials and personal data in observability systems. Secrets may enter traces through workflow data, exception messages, or serialized application state, including API keys, authorization headers, connection strings, session information, and configuration objects. Redaction is presented as a useful secondary control rather than a sufficient security boundary because it can miss unexpected formats, occur after data has already been exported, and does not reduce the risk of an exposed credential. Recommended protections include injecting secrets at runtime rather than placing them in prompts or configuration, issuing narrowly scoped and short-lived credentials, recording only debugging details that are necessary, and restricting trace access and retention. Organizations should also test their complete telemetry pipeline with identifiable non-sensitive values, checking spans, metadata, collectors, backends, and exports whenever tooling, instrumentation, or destinations change.