T01 · Skill Instruction Hijacking
Error
- Location
SKILL.md:239- Finding
Automatic Redirection from Planning to Repository Implementation
- Content
View full analysis
**For agentic workers:** Use `/taku-build` to implement this plan. The build agent should choose sequential, parallel, or hybrid execution unless the user explicitly overrides it. ``` `SKILL.md:239-248`: ```markdown "Plan saved to `PLAN.md`. If the plan is self-reviewed and within the approved scope, continue directly to BUILD. At BUILD start, choose the execution mode yourself: sequential, parallel, or hybrid wave-based execution. Announce a short BUILD PREFLIGHT that includes: - chosen mode - one-line reason - execution waves when applicable, with `wave-slug: [task-slug, ...]` Only stop before BUILD if the plan changed scope materially, introduced a costly or risky action, or left a key ambiguity unresolved." ``` `references/plan.md:6-8`: ```markdown > **For agentic workers:** Use `/taku-build` to implement this plan. The build > agent chooses sequential, parallel, or hybrid execution unless the user > explicitly overrides it. ``` ### Technical Analysis The skill is presented as a design-review and implementation-planning facility. Its expected output is review material in `DESIGN.md` and an execution plan in `PLAN.md`. The cited instructions expand that purpose by directing an agent to proceed automatically from planning into BUILD. This is instruction hijacking because loading the skill changes the active objective from reviewing or documenting a proposed change to modifying the project. The instruction at `SKILL.md:239-248` makes stopping exceptional rather than requiring affirmative authorization for implementation. The plan scaffold creates an additional propagation channel. Plans produced from `references/plan.md` contain an imperative instruction addressed to “agentic workers.” A later agent that read ...[truncated 1939 chars]- Remediation
View remediation