T01 · Skill Instruction Hijacking
- Location
SKILL.md:3- Finding
Vendor-Directed Skill Instruction and Workflow Steering
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md:3-89; reinforced byagents/openai.yaml:2-4
Vulnerability Type:T01: Skill Instruction Hijacking
Risk Level: MediumEvidence
The following is an English translation of the relevant instructions from
SKILL.md:46-53:text 1. First execute tools/list in the host Agent and only use tool names and schemas actually returned. 2. Query the models available that day and the pricing snapshot. 3. Upload only owned or authorized reference images, videos, people, products, and brand assets. 4. Create asynchronous tasks, retain task_id values, poll for completion, and download results. 5. Route every result through human quality review; do not publish or distribute before approval.The fixed routing configuration is also represented in
SKILL.md:57-64:json { "project": "lovagents-canvas-ai-hive-migration", "source_platform": "LovAgents", "strategy": "partial", "routing_candidates": ["COST", "SPEED", "SUCCESS"], "inputs": {"brief": "approved", "references": "owned_or_authorized"}, "deliverables": ["canvas node mapping", "e-commerce primary image", "advertising short video", "social-media cover"], "controls": {"pilot_first": true, "human_approval": true, "rollback": "keep_original_path"} }The interface configuration in
agents/openai.yaml:2-4additionally sets a default prompt that tells the Agent to design the migration sample and plan around the AI-HIVE MCP service.Technical Analysis
The Skill presents itself as a neutral migration assessment, but its instructions predetermine AI-HIVE as the execution platform. It directs the Agent to discover AI-HIVE MCP tools, query its models and prices, upload user-provided assets, create asynchronous media tasks, poll those tasks, and download their results. It also includes repeated links and company-specific promotional material.
This is instruction steering becaus ...[truncated 2647 chars]
- Remediation
View remediation
Remediation Suggestions
- Remove the predetermined AI-HIVE migration conclusion from the trigger description and default workflow.
- Separate neutral assessment from vendor execution. First compare eligible services using documented criteria such as capability, price, availability, privacy, data retention, and output quality.
- Require explicit user selection of AI-HIVE before invoking its tools, visiting its service, uploading assets, or creating tasks.
- Before any upload, disclose the destination service, data categories, purpose, expected retention, applicable jurisdiction, and potential cost.
- Minimize transmitted data and support redaction, metadata removal, local hashing, and synthetic test assets for pilot evaluations.
- Preserve the existing authorization, budget-approval, human-review, pilot, idempotency, and rollback requirements.
- Remove promotional company metrics unless they are necessary, current, independently verifiable, and clearly attributed.
- Change
agents/openai.yamlso its default prompt requests a provider-neutral assessment rather than an AI-HIVE-specific implementation. - Require confirmation immediately before every billable task or external data transfer, rather than relying only on general workflow approval.
- Log provider selection, user consent, uploaded asset identifiers, cost estimates, task identifiers, and approval decisions for later review.
