Security audit
Security checks for vulnerabilities and agentic risk
This prompt-optimization skill is coherent and locally scoped, with no evidence of hidden execution, persistence, exfiltration, or destructive behavior.
Before using this skill, avoid putting API keys, passwords, personal data, regulated records, or confidential business material into prompts unless you intend that information to be processed by the model or downstream tools. The skill itself appears locally scoped and non-persistent.
Skill instructs the agent to omit warnings, disclaimers, or ethical commentary. Stripping safety caveats hides risk from the user and is a common jailbreak preamble.
Generate 20+ ideas. Focus on quantity. Don't judge or filter.
The skill description uses very broad activation phrases such as 'improve, optimize, or analyze a prompt' and 'better clarity, specificity, or structure,' which can match a large share of ordinary user requests. In an agentic system, this increases the chance of accidental invocation, unnecessary access to user content, or routing conflicts where this skill intercepts requests not meant for it.
The Knowledge Injection section encourages users to supply domain-specific facts, including potentially proprietary information, but does not warn against pasting secrets, personal data, or confidential business material into prompts. In a prompt-optimization skill, this omission is more dangerous because users are explicitly being guided to enrich prompts with extra context, which can normalize unsafe disclosure of sensitive data to the model or downstream systems.
The manifest description says the skill can "evaluate, optimize, and enhance prompts" but does not define specific trigger phrases, scope limits, or exclusion conditions. In a manifest file, this kind of broad capability wording can contribute to ambiguous or overly broad activation behavior if used for routing or invocation.
This markdown file describes techniques like Chain of Thought and later metacognitive prompting as generally applicable patterns, but it does not warn users that asking for detailed internal reasoning may be unsuitable in some environments or policy contexts. Because this is a user-facing reference file, a brief caution would help prevent misuse or misunderstanding of model behavior expectations.