Security audit
Security checks for vulnerabilities and agentic risk
The skill is a coherent model-cost advisor, but it asks to persist user preferences, session traces, and learning history without clear opt-in or retention limits.
Install only if you are comfortable with the skill using persistent Memory to personalize model routing and cost reporting. Prefer configuring it so preference learning, session traces, and monthly reports are opt-in, minimized, and deletable.
The trigger rules are overly broad, matching general phrases like saving money, cost concerns, or model choice discussions in ways that can activate the skill outside its intended scope. Unintended activation is dangerous here because the skill then starts preference inference, routing logic, and optional memory behavior on conversations that may not actually require this functionality.
The skill expands from model recommendation into persistent cross-session profiling, storing budget preferences, upgrade history, and other behavioral attributes in memory. That exceeds the narrow functional need of a model-selection advisor and creates avoidable privacy and governance risk if user behavior is tracked without clear necessity, consent, or retention controls.
The long-lived user profile stores strategy, budget, overrides, and upgrade history inferred from conversation content. Even if intended for personalization, this creates persistent user-derived records that can reveal preferences and spending sensitivity, and the skill provides no explicit consent, retention, or minimization safeguards.
The skill gives conflicting instructions about telemetry and logging: one section says to automatically maintain structured session traces after each recommendation/execution, while another says not to proactively update logs every time. In a routed agent system, this ambiguity can lead to inconsistent implementations, unexpected data retention, or accidental over-collection of user interaction data.
The skill instructs the agent to persist detailed per-turn session traces including user intent, router decisions, quality scores, token estimates, cost, and user feedback. This creates a natural-language interaction log that can accumulate sensitive behavioral data and operational metadata well beyond what is needed to answer a model-routing question.
The cross-session learning registry stores historical rejection patterns, task types, confidence labels, and pattern evolution over time. This turns ordinary user interactions into a retained behavioral knowledge base, increasing privacy exposure and creating potential secondary use of user-derived data without strong boundaries.
