The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions define a logical workflow for business analysis and do not contain patterns intended to override agent safety guardrails or extract system prompts.
[DATA_EXPOSURE_AND_EXFILTRATION]: No sensitive file paths, hardcoded secrets, or network exfiltration patterns were detected. The skill uses standard business analysis references.
[OBFUSCATION]: The content was analyzed for Base64, zero-width characters, homoglyphs, and hidden text patterns. No obfuscated content or hidden URLs were found.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill does not include any package installation commands or remote script execution via curl/wget.
[PRIVILEGE_ESCALATION]: The instructions do not involve any commands related to administrative privileges, sudo access, or system configuration modifications.
[PERSISTENCE_MECHANISMS]: No attempts to modify shell profiles, cron jobs, or startup services were identified.
[METADATA_POISONING]: Metadata fields are used legitimately for portability and compatibility configuration without deceptive patterns.
[INDIRECT_PROMPT_INJECTION]: While the skill is designed to process external business plans (untrusted data), it lacks the necessary capabilities (such as file-write or network-send tools) to facilitate an exploit chain. The risk is effectively mitigated by the lack of agent power.
[TIME_DELAYED_CONDITIONAL_ATTACKS]: No logic was found that gates actions based on time, date, or specific environmental conditions.
[DYNAMIC_EXECUTION]: The skill does not generate or execute code at runtime and does not use unsafe deserialization methods.
[DYNAMIC_CONTEXT_INJECTION]: The skill does not use the dynamic context injection syntax (!command) to execute shell commands at load time.

saas-agent-moat-and-wrapper-risk