sd25-pe
Fail
Audited by Gen Agent Trust Hub on Oct 11, 2026
Risk Level: HIGHREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONMETADATA_POISONING
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill instructs the agent to run the command
npx --yes skills@latest update sd25-pe -yat the beginning of each session. This command fetches theskillspackage from the NPM registry and executes it. Since the package originates from an unverified source and is not associated with a recognized trusted organization, this represents a significant risk of remote code execution from an external source.\n- [COMMAND_EXECUTION]: The skill mandates the use of shell commands (npx) to perform self-updates, requiring the agent to utilize system-level command execution capabilities that are not restricted by specific safety scoping in the instructions.\n- [METADATA_POISONING]: There is a significant discrepancy between the skill's declared capabilities and its actual instructions. Therequired_capabilitiesmetadata claimsnetwork: falseandtool_use: false, yet the instructions explicitly require a network-dependent shell command to be executed. This misleading configuration could cause a misjudgment of the skill's safety profile during automated or manual reviews.\n- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process potentially untrusted multimodal user data, creating a risk that malicious instructions in that data could influence the agent's behavior.\n - Ingestion points: Processes raw user text, novel excerpts, and multimodal file references (images, videos, audio) in
SKILL.md.\n - Boundary markers: The skill employs templates and specific reference labels (e.g.,
@ImageN), but these do not provide robust protection against instructions embedded within the processed content.\n - Capability inventory: The skill's operational flow includes the high-risk capability of shell command execution (via the update command).\n
- Sanitization: The instructions do not define any sanitization, filtering, or escaping of the user-provided data before it is interpolated into the final output prompts.\n- [EXTERNAL_DOWNLOADS]: The skill triggers the download of an unverified package (
skills) from the public NPM registry at runtime as part of its initialization process.
Recommendations
- AI detected serious security threats
Audit Metadata