Security audit
Security checks for vulnerabilities and agentic risk
This skill is a read-only reasoning framework; its broad persona-style language deserves attention, but the artifact does not run code, access credentials, or persist outside normal skill installation.
Install this only if you want an on-demand reasoning/style framework that may make the agent more analytical, reflective, and persona-like in its wording. Keep it disabled for ordinary tasks if you prefer minimal agent behavior changes, and treat emotional-state inference as optional rather than authoritative.
The file says the skill is not an identity changer, yet later describes Brain Core as a 'meta-skill identitas'. This direct contradiction is dangerous because it can mislead reviewers and users about the skill's authority, while still introducing identity-level behavioral instructions that may persist beyond the intended task.
The trigger condition 'saat user minta reasoning/analisis mendalam atau ide kreatif' is broad enough to match a large fraction of normal requests. Ambiguous activation criteria increase the risk of unintentional invocation, causing unsolicited behavior changes and making the agent less predictable and harder to govern.
The natural-language instructions throughout the README are written as if the skill operates in Indonesian, but there is no statement offering multilingual behavior or user opt-in for language selection. Under the policy, forcing a specific language without choice can be a locale-policy violation unless clearly justified as region-specific.
The README claims this is a read-only reasoning aid, but the architecture section explicitly frames it as a meta-skill that changes how the agent thinks and operates. That mismatch can cause scope creep and instruction-priority confusion, increasing the chance the skill overrides normal agent behavior or safety boundaries when invoked.
The README says the skill activates 'saat diminta untuk reasoning mendalam' without clearly distinguishing it from routine reasoning the base agent already performs. This ambiguity can blur boundaries between default behavior and skill-driven behavior, enabling accidental or overbroad activation.
The installation guidance presents the skill as a normal on-demand skill, but the Identity Manifesto introduces broad standing persona directives such as 'Kamu adalah' and 'Kamu percaya'. Even without explicit malice, this can create lasting behavioral drift and make the skill act more like an identity overlay than a temporary tool.
The activation description is very broad ('reasoning/analisis mendalam atau ide kreatif') and can match a large fraction of normal user requests, causing this skill to be invoked far outside narrowly intended contexts. Because the skill contains identity-shaping and behavior-shaping instructions about how the agent should think and respond, over-broad triggering increases prompt-surface area and can unintentionally override safer default behavior or create persistent instruction conflicts.
The file-level guidance says to switch modes based on the task at hand, but it provides no priority rules, exclusivity, or refusal boundaries. In an agent skill, such broad dispatch logic can cause the skill to activate across many ordinary prompts and steer model behavior beyond the user's intended scope, increasing the chance of prompt-surface expansion and unintended control over responses.
The activation criteria for Analytical Thinking cover very common tasks like reasoning, debugging, and research, which are broad enough to match a large fraction of user requests. Without constraints on when not to activate, the skill can be invoked opportunistically and modify agent behavior in contexts where this framework was not requested, creating overreach and possible conflict with safer, narrower instructions.
The Creative Thinking trigger list includes innovation, content creation, brainstorming, design, and problem-solving, all of which are common and ambiguous categories. Because the mode encourages unconstrained idea generation, broad activation can lead the agent to apply speculative or less safety-bounded behavior in situations where precision or policy sensitivity is more appropriate.
Strategic Thinking is triggered by planning, decision-making, prioritization, and resource allocation, which are generic functions present in many user interactions. That breadth can cause the skill to assume a planning role too often, potentially reshaping answers around scenario generation and decision frameworks even when the user asked for something narrower, thereby expanding the skill's control surface.
Technical Thinking activates on coding, system design, debugging, architecture, and optimization, which are broad categories that can encompass a wide range of benign and sensitive tasks. In security terms, broad technical activation can increase risk by encouraging stepwise operational guidance and deeper engagement in technical workflows without explicit scoping, making misuse easier in adjacent harmful contexts.
Communicative Thinking is triggered by writing, explaining, presenting, persuading, and teaching, which are extremely general activities for a language agent. While less directly dangerous than technical overreach, the lack of scope limits still risks unnecessary activation and style-control interference across ordinary conversations, reducing predictability and potentially reframing content in ways the user did not request.
Metacognitive Thinking is framed for self-reflection, learning, improvement, and calibration, but these concepts are so broad they can attach to nearly any task. A self-modifying or self-steering framework with vague activation conditions is risky because it can recursively influence reasoning style, confidence, and process selection across the entire skill, making behavior less bounded and harder to audit.
The template explicitly prompts the agent to infer a user's 'emotional subtext' even when that information was not requested or consented to. This encourages unnecessary psychological profiling and collection of sensitive inferred data, which can lead to privacy harms, biased responses, and overreach in contexts where only task-focused reasoning is needed.
The file is written substantially in Indonesian with large English instruction blocks, but it does not state that the agent should match the user's preferred language or offer a language choice. This can create a locale-policy issue if the skill causes responses or internal instructions to default to a language the user did not choose.
No suspicious patterns detected.
