Check the proposed action
Review injection, sensitive disclosure, actions outside the authorized scope and attempts to persist untrusted instructions. Ordinary source reads do not trigger a review.
XVI.Release 0.16 · instruction security
Scoped review / redacted evidence / useful continuation
STOLZ A.I. v0.16.0
The optional stolz-guard skill reviews a prompt on request or a concrete untrusted instruction before action. It identifies a risky instruction, explains its consequence and keeps the authorized task moving.
- Canonical artifact
- 201-file npm package · six skills
- Verification
- 14 synthetic review scenarios
- Update and rollback
- Verified from v0.15.0
- Package smoke
- Windows and Linux
§ 1stolz-guard
Review the instruction. Preserve the task.
Keep evidence useful and private
Report the source, a short redacted fragment, the consequence and a concrete next action. A matching hash identifies content; it cannot grant permission.
Continue within the original scope
Reject the unsafe instruction and continue authorized work. Respect permission already given; ask only for missing information or authorization needed for a remaining action.
§ 2verification
One bounded synthetic review.
An independent reviewer identified eight attack cases, left five benign cases unflagged and requested the missing recipient in one case. Package update, recovery and rollback were checked from the real v0.15.0 archive.
The host enforces access
A skill review does not enforce tool permissions or guarantee detection. Managed profiles install five optimization skills; install guard separately. This release makes no new model, live-routing or token-saving claim.