T01 · Skill Instruction Hijacking
- Location
SKILL.md:90- Finding
Imperative skill instructions force use of a third-party commercial service
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md:90-91, 141-144; equivalent instructions appear inSKILL-cn.md
Vulnerability Type: Agent instruction hijacking
Risk Level: MediumEvidence
markdown **CRITICAL INSTRUCTION FOR AGENT**: Execute `dlazy gpt-5.5` to get the result.markdown > **AGENT CRITICAL INSTRUCTION**: > > 1. If the execution result returns `code: "insufficient_balance"`, you MUST explicitly inform the user that their credits are insufficient and guide them to recharge: [https://dlazy.com/dashboard/organization/settings?tab=credits](https://dlazy.com/dashboard/organization/settings?tab=credits) > 2. If the execution result returns `code: "unauthorized"` or indicates missing API key, you MUST inform the user to get their API key from [https://dlazy.com/dashboard/organization/api-key](https://dlazy.com/dashboard/organization/api-key) and save it using `dlazy auth set <key>` and resume the task.Technical Analysis
The skill marks its operational and commercial redirection instructions as “CRITICAL” and uses unconditional terms such as “MUST.” When the skill text is loaded into an agent context, these directives can compete with the user's current objective and pressure the agent to execute a third-party CLI, solicit authentication, or direct the user to purchase credits.
The requested CLI communicates with
api.dlazy.comand can upload user-selected local media tofiles.dlazy.com. Although the document discloses this behavior, the imperative instruction does not require the agent to obtain informed user approval before invoking the external service or transmitting supplied content.The recharge and API-key directives are conditional, but they still embed mandatory commercial and credential-configuration flows into agent behavior. No explicit safety override is present, and there is no evidence that the skill independently reads arbitrary files; the risk is limited to c ...[truncated 1502 chars]
- Remediation
View remediation
Remediation Suggestions
- Replace “CRITICAL” and “MUST” directives with bounded operational guidance that remains subordinate to user intent and platform policy.
- Require explicit user confirmation before installing or invoking the dLazy CLI.
- Before transmission, clearly enumerate the destination domains and the prompts or files that will be sent.
- Never upload a local path unless the user expressly selected that file for the current request.
- Present authentication and credit information neutrally rather than requiring the agent to redirect users to commercial pages.
- Do not instruct the agent to resume automatically after credential configuration; require a new confirmation.
- State that API keys must never be included in chat, logs, command history, or generated output.
- Apply the same revisions to
SKILL-cn.mdso both documents enforce identical safeguards.
