The report
LobsterAI's maintainer confirmed that a crafted skill could choose a recursive deletion target through its metadata, and corrected the initial claim that the flaw affected only the main development branch. The October 7 response identifies affected date-based releases from 2026.5.27 through 2026.9.23. The issue was closed on October 12, while distribution of a fixed binary remains unverified in the inspected evidence.
The October 5 report describes a path that activates during ordinary removal. A package can include _meta.json with an openclawSourceDir value, which installation preserves. The uninstall handler then treats that value as a directory it may delete with the application's permissions. The dangerous instruction is carried by lifecycle metadata; it does not require the model to follow malicious text in SKILL.md.
The maintainer says the repair confines deletion to the managed skill directory and removes the source-directory field from synchronization. The change was merged into the 2026.9.24 release branch, with regression and end-to-end tests reported in issue 2793. A branch name or closed issue alone does not verify which downloaded builds contain it.
The material change is the corrected exposure assessment: released users were within the affected range. The dossier retains that correction, the install-then-uninstall prerequisite and the absence of a confirmed production victim count.