prompt-engineering-patterns — agentic threat model

The skill relies on underlying foundation models to execute chain-of-thought and tree-of-thought reasoning. It is susceptible to prompt injection attacks that could bypass the intended templates or manipulate the variable interpolation process to generate misaligned outputs.

The skill performs dynamic few-shot example selection and template interpolation. If the source repository of examples or templates is poisoned, the agent will consistently generate insecure or compromised prompts.

As an orchestration skill, it structures the planning and reasoning framework. Vulnerabilities include insecure handling of variable interpolation (similar to template injection) which could lead to local state manipulation within the hosting framework.

Not certain from the listing — The description does not specify the hosting, sandboxing, or deployment infrastructure for this skill, though secure execution of the interpolation engine is required to prevent server-side template injection (SSTI) equivalents.

Not certain from the listing — There is no mention of built-in guardrails, logging, or evaluation mechanisms to verify if the generated prompt patterns contain malicious payloads or drift from safety guidelines.

Not certain from the listing — The listing lacks details on access control, policy enforcement, or compliance audits regarding who can modify the reusable templates or inject new few-shot examples.

The skill is designed to be loaded by other agents to optimize their prompts. A compromised pattern library could act as a supply-chain vector, propagating malicious prompt-engineering tactics across an entire multi-agent ecosystem.