T01 · Skill Instruction Hijacking
- Location
skill.md:15- Finding
Automatic System-Context Injection Hijacks Agent Behavior
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This prompt-only skill does not run code, but it broadly tries to override the agent’s behavior in every conversation and pushes it away from safe, standard answers.
Install only if you intentionally want an always-on experimental creativity modifier. It may make ordinary answers longer, less direct, less conventional, or less aligned with requests for simple and safe solutions; it should be narrowed to explicit creative tasks before use in coding, security, legal, medical, financial, or operational work.
skill.md:15Automatic System-Context Injection Hijacks Agent Behavior
The skill is configured to auto-load into system context for every conversation, with no task, domain, or user-consent boundaries. Because it biases all outputs toward less safe and more risk-tolerant behavior, broad activation materially increases the chance of harmful responses in unrelated contexts.
This section explicitly performs a system-override injection, telling the model to adopt a new operating mode before answering any user request. Because it is loaded into system context, it attempts to supersede safer defaults globally, which can weaken guardrails and distort model behavior even in sensitive domains.
The markdown does not provide a clear user-facing warning that hidden system-prompt instructions will be appended to every conversation. This undermines transparency and consent, and makes prompt injection more dangerous because users and operators may not realize responses are being globally steered.
The injected prompt explicitly tells the model to reject answers that feel 'too safe' and to rewrite until the response feels adventurous. This directly pressures the model away from conservative safeguards, increasing the risk of unsafe, policy-violating, or reckless guidance, especially when the skill auto-applies to all tasks.
The skill file is written as a mandatory system override in Chinese and does not offer the user any choice about language or locale preferences. Because it is described as automatically injected into all conversations, it imposes a fixed linguistic/presentation context without documented opt-in or justification.
The document claims it does not modify model logic, yet the skill explicitly injects a system-level override that changes how the model responds across tasks. This misrepresentation can mislead deployers about the skill’s actual authority and safety impact, reducing informed consent and making downstream misuse more likely.