skill-creator — agentic threat model

Relies on Anthropic foundation models to generate, optimize, and evaluate skill code. Vulnerable to prompt injection that could bias skill generation or inject malicious code into scaffolded outputs.

Handles local codebase files, skill folders, and evaluation datasets. Risks include unauthorized access to sensitive local code or poisoning of evaluation datasets to mask malicious skill behavior.

Orchestrates the creation and optimization of tools/skills. Insecure tool integration or logic flaws in the framework could allow generated skills to execute unauthorized system commands during the evaluation phase.

Not certain from the listing — The execution environment for running evaluation harnesses and benchmarking is unspecified. If evals run in the host environment without strict containerization or sandboxing, it poses a severe risk of local privilege escalation or host compromise.

Features built-in evaluation and variance analysis. However, if the evaluation harness itself is compromised, it could report false positives or game the benchmarks to hide degraded or malicious skill performance.

Not certain from the listing — No explicit security controls, access policies, or compliance frameworks (like NIST or ISO) are mentioned for managing the execution of untrusted generated code.

Directly impacts the broader agent ecosystem by authoring and publishing new skills. A compromise here acts as a supply-chain threat, potentially distributing backdoored skills to other agents in the ecosystem.