Dbs Agent Migration Skill Security Audit

The risk is supported. The workflow actively calls for deleting rules the model judges obsolete, duplicated, or too host-specific; a mistaken judgment could remove safeguards, project conventions, or required host settings. It must disclose proposed deletions before writing, which adds visibility, but the source does not require a backup, item-by-item approval, a retained diff, or post-deletion behavioral tests. Users can require a full diff, approve each deletion, and keep a recoverable copy.

The risk is supported, though the operation requires explicit permission. The workflow may repoint an existing same-name symlink, and the global directory is read by Doubao, Trae Solo, and Codex; once authorized, several hosts may therefore load a different Skill. It requires permission before writing and verification with readlink, but does not require listing every affected application or preserving the old target. Users can authorize only project-local bridges and request a record of the prior target and affected hosts.

The risk is plausible, but this does not show an attack occurred. The workflow selects project Skills as sources and creates bridges that tell hosts to read and follow them. Its filtering only excludes articles, backups, tests, and exports; it does not require checks for credential uploads, destructive file actions, or external instructions. User confirmation reduces accidental changes but does not detect a malicious source. Users can ask for a per-source security review and restrict bridges to reviewed files.