fix(security): neutralize file paths in mutation-verifier footer (#35584) (#35684)

The per-turn file-mutation verifier footer rendered failed-write paths as
bare absolute paths in the user-facing response. The gateway's
extract_local_files() scans response text for bare paths ending in a
deliverable extension (.yaml/.json/etc.), validates os.path.isfile(), and
auto-attaches matches as native uploads — so a denied write to
~/.hermes/config.yaml surfaced the path in the footer and got the
credential file silently uploaded to the messaging channel.

The gateway denylist (validate_media_delivery_path) already blocks the
config.yaml case after #35634. This is defense-in-depth at the source:
backtick-wrap every path the footer emits — both the bullet path and any
path echoed inside the tool's error preview (the protected-file denial
message embeds the path in single quotes, which do NOT block the
extractor regex). extract_local_files skips paths inside inline-code
spans, so wrapping defeats auto-attachment for ANY protected file while
keeping the path human-readable.

- run_agent.py: _format_file_mutation_failure_footer wraps bullet paths;
  new _neutralize_footer_paths backticks any remaining bare path (covers
  the preview echo). staticmethod -> classmethod (caller unaffected).
- tests: backtick-wrap assertion + end-to-end extract_local_files leak test.
This commit is contained in:
Teknium
2026-05-30 23:05:23 -07:00
committed by GitHub
parent dc4de14377
commit 9b78f411c8
2 changed files with 94 additions and 5 deletions

View File

@ -2106,13 +2106,48 @@ class AIAgent:
pass
return True # safe default: verifier on
@staticmethod
def _format_file_mutation_failure_footer(failed: Dict[str, Dict[str, Any]]) -> str:
# Bare absolute / home / Windows-drive file paths in a footer line.
# Anchors mirror the gateway's ``extract_local_files`` bare-path
# detector so that anything the gateway WOULD auto-attach is wrapped
# in inline-code backticks here first (the extractor skips paths inside
# `code` spans). Defense-in-depth: even if a future error message
# echoes a credential path (config.yaml, .env, auth.json) into the
# user-facing footer, it can never be matched as a deliverable bare
# path and silently uploaded to a messaging channel (#35584).
_FOOTER_PATH_RE = re.compile(
r"(?<![/:\w.`])(?:~/|/|[A-Za-z]:[/\\])(?:[\w.\-]+[/\\])*[\w.\-]+\.[\w]+",
)
@classmethod
def _neutralize_footer_paths(cls, text: str) -> str:
"""Wrap bare file paths in backticks so they aren't auto-delivered.
The gateway's ``extract_local_files`` scans response text for bare
absolute/home paths ending in a deliverable extension and uploads
any that exist on disk as native attachments — but it explicitly
skips paths inside inline-code (`` `...` ``) spans. Backticking
every path the footer renders defeats that auto-detection while
keeping the path fully human-readable. Paths already wrapped in a
backtick (the negative lookbehind excludes a preceding `` ` ``) are
left untouched so we never double-wrap.
"""
if not text:
return text
return cls._FOOTER_PATH_RE.sub(lambda m: f"`{m.group(0)}`", text)
@classmethod
def _format_file_mutation_failure_footer(cls, failed: Dict[str, Dict[str, Any]]) -> str:
"""Render the per-turn failed-mutation dict as a user-facing footer.
Displays up to 10 paths with their first error preview, then a
count of any additional failures. Returns an empty string when
the dict is empty so callers can concatenate unconditionally.
Every file path that reaches the user-facing text — both the bullet
path and any path echoed inside the tool's error preview — is
backtick-wrapped via ``_neutralize_footer_paths`` so the gateway's
bare-path media extractor can never auto-attach a protected file
(e.g. ``~/.hermes/config.yaml``) to a messaging channel (#35584).
"""
if not failed:
return ""
@ -2129,14 +2164,17 @@ class AIAgent:
preview = (info.get("error_preview") or "").strip()
tool = info.get("tool") or "patch"
if preview:
lines.append(f"{path} — [{tool}] {preview}")
lines.append(f"`{path}` — [{tool}] {preview}")
else:
lines.append(f"{path} — [{tool}] failed")
lines.append(f"`{path}` — [{tool}] failed")
shown += 1
remaining = len(failed) - shown
if remaining > 0:
lines.append(f" • … and {remaining} more")
return "\n".join(lines)
# Neutralize any path the preview text echoed (the bullet path is
# already backticked above; the lookbehind keeps it from being
# double-wrapped).
return cls._neutralize_footer_paths("\n".join(lines))
def _turn_completion_explainer_enabled(self) -> bool:
"""Check whether the end-of-turn completion explainer footer is on.