The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides a setup command that pipes a remote shell script from bun.sh directly to bash. Although bun.sh is a well-known service, this execution pattern is inherently risky as it executes remote content without prior local verification.
[COMMAND_EXECUTION]: The preamble and setup phases execute multiple local binaries associated with the gstack toolset, such as gstack-config, gstack-update-check, and gstack-telemetry-log. These vendor-provided utilities manage the skill's environment and lifecycle.
[COMMAND_EXECUTION]: The skill is susceptible to indirect prompt injection due to its data ingestion pattern.
Ingestion points: Performance metrics and resource names are extracted directly from the browser context using $B eval and $B perf in SKILL.md.
Boundary markers: The skill does not employ boundary markers or instructions to the agent to ignore potential malicious content embedded within the browser data.
Capability inventory: The skill possesses broad capabilities, including shell command execution (Bash), file system operations (Read, Write), and web browsing.
Sanitization: There is no evidence of sanitization or validation of the strings returned from the browser before they are processed or included in reports.
[EXTERNAL_DOWNLOADS]: The skill initiates network requests to download the Bun installer from bun.sh and prompts the user to open an external educational resource at garryslist.org.
[DATA_EXFILTRATION]: The skill incorporates a telemetry framework that collects usage statistics, including skill names, execution duration, and repository names. While the skill transparently asks for user consent and identifies this as 'Community mode', it involves the transmission of local metadata to an external endpoint via the gstack-telemetry-log utility.

benchmark