Security audit
Security checks for vulnerabilities and agentic risk
This skill has little technical access, but it is designed and documented to generate shareable insults about real people with broad auto-invocation and limited safeguards.
Review this skill carefully before installing. It does not appear to access private files, credentials, or the network, but it is intentionally built to produce harsh insults for social sharing and may be inappropriate for real people, workplace contexts, minors, vulnerable users, or any harassment scenario.
The skill explicitly encourages generating insulting replies about other people and emphasizes screenshot-sharing and social media use, which increases the likelihood of harassment, bullying, and reputational harm. The context makes this more dangerous because the intended use cases target real people in online interactions without any safeguards, consent checks, or warnings about abuse.
Mandating an 'internet troll' style as the only output mode hardcodes aggressive behavior and removes safer alternatives, making harmful or abusive outputs more likely by design. In this skill's context, that risk is amplified because the workflow is specifically optimized to produce sharable insults from arbitrary user-provided text.
The description is broadly framed as a general-purpose 'internet troll roaster' with variable output behavior, but it does not define clear activation boundaries, safety constraints, or disallowed targets. In an agent ecosystem, this can cause the skill to be invoked too readily for harassment, abusive content generation, or escalations against real individuals, increasing the chance of policy-violating or harmful output.
The instruction says the system will 'automatically call' the skill when text is input directly, without any clear trigger boundaries, consent, or scoping. Overly broad auto-invocation can cause this harassment-oriented skill to activate on arbitrary user content, increasing the chance of unintended abusive output, policy bypass via routing, or misuse in contexts where insulting responses are unsafe or inappropriate.
The Chinese instructions repeat the same broad behavior: direct text input causes automatic execution with no explicit boundaries. Because the skill is designed to produce insulting content, ambiguous auto-triggering makes accidental activation and harmful outputs more likely across multilingual contexts, widening exposure rather than limiting it.
The prompt construction explicitly sets a default English reply and switches to Chinese only when Chinese characters are detected. This imposes a language-selection policy on the user without offering a choice or documenting a justified locale restriction, which matches the language/locale policy violation criterion.
The docstring at L103-L105 says the function generates roast content based on input language, and the code builds a prompt at L106, implying input-driven generation. In reality, the function only uses language detection to choose between two hardcoded responses and never uses the constructed prompt, which contradicts the stated behavior of generating roast content from the input.
No suspicious patterns detected.
