fix(profiles): cross-profile soft guard on file-write tools + system-prompt hint (#31290)

* fix(profiles): cross-profile soft guard on file-write tools + system-prompt hint

Adds a soft guard so an agent running under one Hermes profile cannot
silently edit a different profile's skills/plugins/cron/memories.
Three layers:

A. agent/file_safety.classify_cross_profile_target
   Classifies a write target against the active HERMES_HOME. Returns
   a {active_profile, target_profile, area, target_path} dict when the
   path lands in another profile's scoped area. PROFILE_SCOPED_AREAS =
   (skills, plugins, cron, memories). get_cross_profile_warning()
   wraps it into a model-facing error string that names both profiles,
   names the area, and points at the cross_profile=True bypass.

   Defense-in-depth, NOT a security boundary — the terminal tool runs
   as the same OS user and can write any of these paths directly. The
   guard exists to prevent confused-agent corruption, not to stop a
   determined attacker. SECURITY.md §3.2 (terminal-bypass posture)
   still applies.

   Wired into tools/file_tools.write_file_tool and patch_tool with a
   cross_profile=False kwarg. WRITE_FILE_SCHEMA and PATCH_SCHEMA both
   advertise cross_profile so the model can pass it after explicit
   user direction. patch_tool extracts target paths from V4A patch
   bodies before checking (same shape as the existing sensitive-path
   check).

   skill_manage is already scoped to the active profile's SKILLS_DIR
   by construction, so no extra guard wiring is needed there. The
   D-side error message (below) still names other profiles when the
   skill exists elsewhere.

B. agent/system_prompt
   One deterministic line near the environment-hints block names the
   active profile and tells the model not to modify another profile's
   skills/plugins/cron/memories without explicit direction. Profile
   name is stable for the lifetime of the AIAgent, so the line is
   prompt-cache-safe.

D. tools/skill_manager_tool._skill_not_found_error
   Replaces the bare "Skill 'X' not found." with a message that:
     - names the active profile,
     - searches OTHER profiles' skills dirs for the same name,
     - names the profile(s) where the skill exists and the path,
     - suggests `hermes -p <name>` to switch profiles, or
       cross_profile=True for an explicit edit.

   All 5 "not found" sites in skill_manager_tool (edit, patch, delete,
   write_file, remove_file) now go through the helper.

Reference incident (May 2026): a hermes-security profile session
edited skills under both ~/.hermes/profiles/hermes-security/skills/
AND ~/.hermes/skills/ (the default profile's skills) without
realizing the second path belonged to a different profile. Three of
the four skill files needed manual restoration afterward.

What this PR does NOT do:

  * No hard block. The terminal tool can still touch any of these
    paths with no guard — same posture as the dangerous-command
    approval flow. SECURITY.md §3.2 applies.
  * No regex sweep on terminal commands for cross-profile paths.
    That direction is a Skills-Guard-style arms race (cd + relative
    paths, base64, etc.) and would false-positive on legitimate
    cross-profile reads. Filed as a follow-up.
  * No on-disk path migration. ~/.hermes/skills/ remains the
    default profile's skills dir; this PR is about telling the
    agent about that boundary, not changing the layout.

Tests:
  tests/agent/test_file_safety_cross_profile.py (16 tests)
    - _resolve_active_profile_name covers default/named/failure paths
    - classify_cross_profile_target covers all four scoped areas,
      both directions (default → named, named → default, named → named),
      non-Hermes paths, and root-level config files
    - get_cross_profile_warning covers in-profile no-op, cross-profile
      message shape, and the defense-in-depth self-documentation

  tests/tools/test_cross_profile_guard.py (12 tests)
    - write_file: in-profile allow, cross-profile block, cross_profile=True
      bypass, non-Hermes pass-through
    - patch: replace-mode block, cross_profile=True bypass, V4A patch
      path extraction
    - skill_manage: error names the other profile (single + multiple),
      missing-everywhere falls back to skills_list hint
    - system prompt: contract-level checks (both branches present,
      cross_profile=True mentioned, ~/.hermes/profiles/ referenced)

All 207 existing tests in file_safety/file_operations/skill_manager
still pass. 10 system-prompt tests still pass.

E2E verified: the exact incident scenario (security profile editing
default's hermes-agent-dev skill) is now blocked with the warning
message; cross_profile=True unblocks.

* fix(code_execution): add cross_profile to write_file/patch stubs

The cross_profile kwarg added to write_file_tool/patch_tool needs to
flow through the execute_code sandbox stubs in _TOOL_STUBS so the
test_stubs_cover_all_schema_params drift test passes. Without this,
scripts running inside execute_code couldn't pass cross_profile=True
through hermes_tools.write_file().

Caught by CI on PR #31290.
This commit is contained in:
Teknium 2026-05-24 00:38:17 -07:00 committed by GitHub
parent b207dc28b3
commit d3c167b644
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
7 changed files with 846 additions and 19 deletions

View file

@ -174,6 +174,37 @@ def _check_sensitive_path(filepath: str, task_id: str = "default") -> str | None
return None
def _check_cross_profile_path(filepath: str, task_id: str = "default") -> str | None:
"""Return a cross-profile warning string when ``filepath`` lands in
another Hermes profile's skills/plugins/cron/memories directory.
Returns ``None`` when the write is in-scope (same profile) or outside
Hermes scope entirely. Soft guard the agent can override by passing
``cross_profile=True`` to its write tool after explicit user direction.
Defense-in-depth, NOT a security boundary the terminal tool runs
as the same OS user and can write any of these paths directly.
See ``agent/file_safety.classify_cross_profile_target`` for the
detection rules.
"""
try:
from agent.file_safety import get_cross_profile_warning
except Exception:
# Fail open on import error — the existing sensitive-path guard
# plus the write_denied list still apply.
return None
# Resolve via the task's cwd so a relative ``skills/foo/SKILL.md``
# in a session that cd'd into ``~/.hermes/profiles/other/`` is
# classified against the right base.
try:
resolved = str(_resolve_path_for_task(filepath, task_id))
except (OSError, ValueError):
resolved = filepath
return get_cross_profile_warning(resolved)
def _is_expected_write_exception(exc: Exception) -> bool:
"""Return True for expected write denials that should not hit error logs."""
if isinstance(exc, PermissionError):
@ -795,11 +826,23 @@ def _check_file_staleness(filepath: str, task_id: str) -> str | None:
return None
def write_file_tool(path: str, content: str, task_id: str = "default") -> str:
"""Write content to a file."""
def write_file_tool(path: str, content: str, task_id: str = "default",
cross_profile: bool = False) -> str:
"""Write content to a file.
``cross_profile`` opts out of the soft cross-Hermes-profile guard. The
guard fires only on writes that land in another profile's
skills/plugins/cron/memories directory; everything else is unaffected.
Pass ``True`` after explicit user direction same shape as ``force``
on the terminal tool.
"""
sensitive_err = _check_sensitive_path(path, task_id)
if sensitive_err:
return tool_error(sensitive_err)
if not cross_profile:
cross_warning = _check_cross_profile_path(path, task_id)
if cross_warning:
return tool_error(cross_warning)
if _is_internal_file_status_text(content):
return tool_error(
"Refusing to write internal read_file status text as file content. "
@ -854,8 +897,13 @@ def write_file_tool(path: str, content: str, task_id: str = "default") -> str:
def patch_tool(mode: str = "replace", path: str = None, old_string: str = None,
new_string: str = None, replace_all: bool = False, patch: str = None,
task_id: str = "default") -> str:
"""Patch a file using replace mode or V4A patch format."""
task_id: str = "default", cross_profile: bool = False) -> str:
"""Patch a file using replace mode or V4A patch format.
``cross_profile`` opts out of the soft cross-Hermes-profile guard for
targets under another profile's skills/plugins/cron/memories
directory. Same shape as ``write_file``'s flag.
"""
# Check sensitive paths for both replace (explicit path) and V4A patch (extract paths)
_paths_to_check = []
if path:
@ -868,6 +916,10 @@ def patch_tool(mode: str = "replace", path: str = None, old_string: str = None,
sensitive_err = _check_sensitive_path(_p, task_id)
if sensitive_err:
return tool_error(sensitive_err)
if not cross_profile:
cross_warning = _check_cross_profile_path(_p, task_id)
if cross_warning:
return tool_error(cross_warning)
try:
# Resolve paths for locking. Ordered + deduplicated so concurrent
# callers lock in the same order — prevents deadlock on overlapping
@ -1052,7 +1104,12 @@ WRITE_FILE_SCHEMA = {
"type": "object",
"properties": {
"path": {"type": "string", "description": "Path to the file to write (will be created if it doesn't exist, overwritten if it does)"},
"content": {"type": "string", "description": "Complete content to write to the file"}
"content": {"type": "string", "description": "Complete content to write to the file"},
"cross_profile": {
"type": "boolean",
"description": "Opt out of the cross-profile soft guard. Defaults to false. Set true ONLY after explicit user direction to edit another Hermes profile's skills/plugins/cron/memories — by default these writes are blocked with a warning because they affect a different profile than the one this session is running under.",
"default": False,
},
},
"required": ["path", "content"]
}
@ -1099,6 +1156,11 @@ PATCH_SCHEMA = {
"type": "string",
"description": "REQUIRED when mode='patch'. V4A format patch content. Format:\n*** Begin Patch\n*** Update File: path/to/file\n@@ context hint @@\n context line\n-removed line\n+added line\n*** End Patch",
},
"cross_profile": {
"type": "boolean",
"description": "Opt out of the cross-profile soft guard. Defaults to false. Set true ONLY after explicit user direction to edit another Hermes profile's skills/plugins/cron/memories.",
"default": False,
},
},
"required": ["mode"],
},
@ -1149,7 +1211,10 @@ def _handle_write_file(args, **kw):
f"write_file: 'content' must be a string, got "
f"{type(args['content']).__name__}."
)
return write_file_tool(path=args["path"], content=args["content"], task_id=tid)
return write_file_tool(
path=args["path"], content=args["content"], task_id=tid,
cross_profile=bool(args.get("cross_profile", False)),
)
def _handle_patch(args, **kw):
@ -1157,7 +1222,9 @@ def _handle_patch(args, **kw):
return patch_tool(
mode=args.get("mode", "replace"), path=args.get("path"),
old_string=args.get("old_string"), new_string=args.get("new_string"),
replace_all=args.get("replace_all", False), patch=args.get("patch"), task_id=tid)
replace_all=args.get("replace_all", False), patch=args.get("patch"), task_id=tid,
cross_profile=bool(args.get("cross_profile", False)),
)
def _handle_search_files(args, **kw):