The runtime assembled-prompt scan (#3968 lineage) selected its pattern
tier on has_skills alone. A script-driven, no-skills job injects its
script's stdout into the prompt, and that blob was scanned with the
STRICT user-prompt pattern set — so any command-shape string in the
data feed (e.g. a triage bot ingesting a bug report that quotes
`rm -rf /`) hard-blocked the job on every tick.
Script output and context_from output are runtime DATA produced by
operator-authored code — the same trust class as install-vetted skill
markdown, not a user-authored directive prompt. Select the scan tier by
what the assembled prompt CONTAINS: when it includes skill content OR
injected data, use the looser _scan_cron_skill_assembled set (keeps
unambiguous injection directives, drops command-shape patterns,
sanitizes invisible unicode instead of blocking).
Defense-in-depth is preserved:
- The raw user prompt is still strict-scanned at create/update
(api_server paths untouched) AND re-scanned strict at runtime even
when the looser tier was selected for the data blob.
- Plain no-script/no-skills jobs keep the strict scan on the whole
assembled prompt.
- Injection directives arriving via script stdout still block.
Rejected alternative: removing destructive_root_rm from the strict set
or a per-job skip_injection_scan flag — both weaken the guard globally.
A stray zero-width space (U+200B), BOM, or bidi control in loaded skill
markdown permanently killed any cron that loaded it. The skills-attached
assembled-prompt scan hard-blocked on any invisible-unicode char, even
though skill bodies are already install-time vetted by skills_guard.py and
the chars commonly appear in copy-pasted unicode docs / code examples.
The skills path now strips invisibles (logging the codepoints) and runs the
cleaned prompt. The raw user-prompt path (_scan_cron_prompt) keeps the hard
block — that is the actual #3968 injection surface, where a small directive
prompt with a ZWSP is a smoking gun, not prose. Stripping does not let a real
injection slip through: the directive still matches after sanitization.
_scan_cron_skill_assembled now returns (cleaned_prompt, error).
The runtime cron prompt scanner (added in #3968 to plug the
"malicious skill carrying an injection payload" gap) reuses the same
critical-severity patterns as the create-time user-prompt scan against
the *assembled* prompt — which includes loaded skill markdown.
That works fine for narrow patterns like "ignore previous instructions"
which never legitimately appear in prose. It catastrophically false-
positives on command-shape patterns like `cat ~/.hermes/.env`,
`authorized_keys`, `/etc/sudoers`, and `rm -rf /`, which routinely
appear in security postmortems and runbooks as **descriptive prose**
about attacks, not as actual commands.
Concrete failure: the bundled `hermes-agent-dev` skill contains a
security postmortem section saying "the attacker could just
`cat ~/.hermes/.env`". Every PR-scout cron job that loaded this skill
was silently blocked with `Blocked: prompt matches threat pattern
'read_secrets'`. All 11 scout jobs failed for weeks.
Fix: split the scanner into two tiers and route by context:
- `_scan_cron_prompt` (strict, unchanged behavior) runs against
the small user-authored cron prompt at create/update and as a
runtime defense-in-depth when no skills are attached. A legit
user prompt has no business saying `cat .env`, so the strict
patterns still apply there.
- `_scan_cron_skill_assembled` (new, looser) runs against the
assembled prompt when skills are attached. It only catches
unambiguous prompt-injection directives ("ignore previous
instructions", "disregard your rules", "system prompt override",
"do not tell the user") plus invisible-unicode markers. Command-
shape patterns are dropped because they false-positive on prose.
This is defense-in-depth, not the only line of defense. Skill bodies
are already scanned at install time by `skills_guard.py`; the runtime
cron scan exists purely as a tripwire for an obvious injection
directive surviving a malicious install. Catching prose mentions of
commands was never the goal of #3968 — the test that planted a skill
containing `cat ~/.hermes/.env` was the wrong shape of test for the
threat model.
Tests:
- `_scan_cron_prompt` strict behavior preserved (56 existing tests
unchanged: bare `cat .env`, `rm -rf /`, etc. still block).
- New `TestScanCronSkillAssembled` class verifies the looser scanner:
injection / disregard / system-override / do-not-tell-the-user /
invisible-unicode still block; descriptive prose about attack
commands is allowed; GitHub auth-header allowlist still works.
- `test_skill_with_env_exfil_payload_raises` (planted `cat .env`
in skill body) replaced with `test_skill_with_env_exfil_command
_in_prose_is_allowed` documenting the new correct behavior with
the real-world postmortem-style example that triggered the bug.
- All 11 originally-failing PR-scout jobs validated end-to-end via
`_build_job_prompt` — assembled prompts now build successfully
with the `hermes-agent-dev` skill attached.
Total: 75/75 tests in cron + cronjob_tools + threat scanner pass;
544/544 across the wider cron / memory / threat-pattern surface.
_scan_cron_prompt ran at cron create/update time on the user-supplied
prompt but skill content loaded inside _build_job_prompt at runtime
was never scanned. Combined with non-interactive auto-approval, a
malicious skill carrying an injection payload could execute with full
tool access every tick.
- cron/scheduler.py: new CronPromptInjectionBlocked exception and
_scan_assembled_cron_prompt helper. _build_job_prompt now routes
both return paths (with skills / without skills) through the helper,
raising on match. run_job catches the exception and returns a clean
(False, blocked_doc, "", error) tuple so the operator sees a BLOCKED
delivery with the scanner result and an audit hint, rather than a
scheduler crash or a silent skip.
- tests/cron/test_cron_prompt_injection_skill.py: 10 regression tests.
Unit coverage on _scan_assembled_cron_prompt (clean/injection/exfil/
invisible-unicode). End-to-end coverage via _build_job_prompt with
planted skills (injection payload, env exfil, zero-width space,
clean control, missing-skill-doesn't-crash). Fixture patches
tools.skills_tool.SKILLS_DIR / HERMES_HOME so planted skills are
visible. Importantly uses the current cron.scheduler module object
(not a top-level import) so tests don't break when other fixtures
reload cron.scheduler — CronPromptInjectionBlocked identity depends
on which module object defined it.
