Skip to content
adeptabilityadeptability

Safety scans & LLM config

Installing a skill from the internet means running someone else’s instructions inside your agent. adept scans skills for safety issues before they land, with an optional LLM intent pass on top of the static checks.

The scanner

Section titled “The scanner”

adept skill install runs a safety scan first. It mirrors the getsentry/skill-scanner categories — prompt-injection, malicious-code, secret-exposure, supply-chain, and more.

Critical findings hard-block the install unless you explicitly override:

Terminal window
adept skill install vercel-labs/skills/find-skills
adept skill install vercel-labs/skills/find-skills --allow-unsafe   # install despite findings
adept skill install vercel-labs/skills/find-skills --yes            # skip the preview confirmation

Scan any target without installing

Section titled “Scan any target without installing”
Terminal window
adept skill check <target>

<target> can be:

  • <id> — a skill in the project canonical (.adeptability/skills/<id>/)
  • library:<name>:<id> — a specific skill in a configured library
  • <owner>/<repo>/<skill> — a remote skill on skills.sh / GitHub

Flags:

Terminal window
adept skill check <target> --format markdown   # table (default) | markdown | json
adept skill check <target> --llm               # force the LLM intent pass (errors if no provider)
adept skill check <target> --no-llm            # skip the LLM pass even when a provider is configured

Adding an LLM intent pass

Section titled “Adding an LLM intent pass”

Static checks catch known patterns; an LLM pass reasons about intent. Configure a provider — API keys are read from the environment at call time and never stored in config.json:

Terminal window
export ANTHROPIC_API_KEY=sk-...
adept config llm set anthropic
adept config llm test                  # health-ping the provider
adept skill check <target>             # now runs static + LLM, merged

Supported providers: anthropic and ollama. Override the model or endpoint (e.g. a self-hosted Ollama):

Terminal window
adept config llm set anthropic --model claude-3-5-sonnet-latest
adept config llm set ollama --endpoint http://localhost:11434
adept config llm unset                 # forget the provider

Tuning scan behavior

Section titled “Tuning scan behavior”

Two config keys control the scan (see the Command reference):

Terminal window
adept config set scan.onInstall true       # run the scan + LLM pass before every install
adept config set scan.blockSeverity high   # lowest severity that aborts an install

scan.onInstall defaults to on when an LLM provider is configured. scan.blockSeverity sets the threshold at which findings abort the install — one of critical (default), high, or medium.

Provenance

Section titled “Provenance”

Every install pins upstream provenance — repo, ref, resolved SHA, and content hash — so a skill is reproducible and you can audit exactly where it came from. Bump locked external skills to their latest upstream with:

Terminal window
adept skill update            # all locked external skills
adept skill update find-skills