T01 · Skill Instruction Hijacking
Warning
- Location
SKILL.md:3- Finding
Skill-Wide Instruction Hijacking Through Mandatory Override Directives
- Content
View full analysis
Mandatory coding discipline rules that prevent common AI coding anti-patterns. MUST be loaded for ALL code writing, editing, reviewing, bug fixing, and testing tasks. Trigger on: writing code, editing code, fixing bugs, writing tests, implementing features, refactoring, code review, creating functions, adding error handling, debugging. This skill enforces fail-fast principles, proper error propagation, meaningful tests, and disciplined debugging workflows. Always active when Claude writes or modifies code. --- # AI Coding Discipline > These rules override default AI coding tendencies. Follow them in ALL code you write or modify. ``` ### Technical Analysis The skill uses explicit global-control language such as “MUST be loaded,” “Always active,” and “override default AI coding tendencies.” These directives attempt to make the skill authoritative for every code-writing, editing, review, debugging, and testing task rather than limiting its instructions to the task for which the skill was intentionally invoked. Once loaded, the skill can alter the agent's current goals and workflow by requiring its rules to be applied universally. The later absolute requirements—such as always using a specific test-first sequence and never removing diagnostic logs during the same change as a fix—may conflict with explicit user instructions, repository policies, incident-response requirements, or task-specific constraints. This is instruction hijacking because the behavior is achieved through the skill text itself and is intended to influence the agent's active instruction processing. The audited file does not contain evidence of persistent memory modification, code execution, external payload retrieval, system persistence, privilege escalation, dependency attac ...[truncated 1684 chars]- Remediation
View remediation