The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides instructions to download and execute scripts directly from remote URLs, a practice that bypasses standard package verification.
In references/playbooks/hermes-agent.md, the instruction curl -fsSL https://get.hermes.dev | sh executes an unverified remote script.
In references/playbooks/lm-studio.md, the instruction curl -fsSL https://lmstudio.ai/install.sh | sh executes an installer from a third-party domain.
In references/playbooks/openclaw.md, the instruction curl -fsSL https://get.openclaw.ai | sh downloads and runs a script directly.
In scripts/pxe-provisioning.md, a script is downloaded from https://raw.githubusercontent.com/apt-mirror/apt-mirror/master/apt-mirror, given executable permissions, and run with root privileges via sudo.
[EXTERNAL_DOWNLOADS]: The skill heavily relies on external software, models, and configurations from numerous non-whitelisted sources.
Frequent downloads from github.com, gitlab.com, and various project-specific domains for tools and AI models.
Automated installation of dependencies from third-party repositories without integrity checks.
[COMMAND_EXECUTION]: The documentation frequently suggests high-privilege operations that significantly reduce the system's security posture.
Instructions in references/support/field-diagnostic-software.md and scripts/recovery.md require the user to disable Secure Boot in the UEFI firmware.
In scripts/pxe-provisioning.md, the skill instructs users to configure fwupd to allow untrusted binaries by setting OnlyTrusted=false in fwupd.conf.
Use of the dd command for disk writing (scripts/pxe-provisioning.md) presents a risk of accidental data destruction.
[DATA_EXFILTRATION]: While no active exfiltration was detected, the skill directs users to handle sensitive secrets and credentials, creating a risk of exposure.
Users are guided to manage API keys for services like NVIDIA, HuggingFace, and Tavily, and to create .env files which could be targeted by malicious scripts.

dgx-spark