Evidence record / hermes-agent
A dated record of one change, kept so the writing that cites it can be checked. Compiled from the sources listed below by the research run, not written for reading. The judgment lives in the issues that cite it, below.
2026-08-10-hermes-agent-terminal-exception-results-and-acp-stderr-were-handing-raw-secrets-to
Terminal exception results and ACP stderr were handing raw secrets to the model.
Both terminal_tool exception paths (generic except and TERMINAL_DEGRADED_MODE=fail) returned raw str(e) plus traceback.format_exc() to the model -- only the logger copy was redacted, so a failing command with an inline secret put that secret in the model's context and in the transcript. Separately, the ACP adapter's _setup_logging cleared root handlers and installed a plain logging.Formatter, bypassing redaction on ACP stderr entirely. Both now pass through redact_sensitive_text / RedactingFormatter. The before/after table shows OPENAI_API_KEY=sk-proj-AbCd... verbatim becoming OPENAI_API_KEY=***.
Channel: tagged-release. Ancestry: merge_commit_sha 72eda946be949a5932923df3325037a0d6c5da49; compare/72eda946...v2026.8.13 -> status=ahead, ahead_by=996, behind_by=0 (ancestor of stable tag v2026.8.13).
Operator consequence: This is a transcript-contamination class, not a live-exploit class: the secret is in stored session data and in anything you exported or shared from it. If you ran commands with inline credentials on v0.20.0 or earlier, rotate them and treat old session exports as sensitive.
Receipt
Finding metadata
Run: 2026-08-10-weekly-digest-2026-08-03_2026-08-10-frontier-v0
Finding ID: 2026-08-10-hermes-agent-terminal-exception-results-and-acp-stderr-were-handing-raw-secrets-to
Source links
Primary links, including exact changelog lines when available.
Versioned source: run artifact