Agent Skill · Anthropic
Word (docx) skill
Anthropic's published skill for reading and generating Word documents with correct structure, styles and tracked-change handling.
This skill bundles executable code. Scripts shipped alongside the instructions run with the permissions of the engineer running the agent. Review the scripts, not only the SKILL.md.
What it is, and what it costs you
Producing a Word document that opens correctly and respects a house style is a structural task, not a writing task. The `.docx` format is a zipped XML package where styles, numbering, headers and revision marks all live in specific places, and a model emitting XML by hand gets it subtly wrong in ways that only surface when someone opens the file. This skill delegates the structure to code and keeps the model on the content.
For a security team the interesting property of `.docx` is how much it carries that is not the visible text. Tracked changes and comments preserve deleted content — including the paragraph somebody removed precisely because it should not be shared. Document properties retain authors, revision counts, and sometimes file paths. An agent reading a document reads all of it, and an agent *generating* one can propagate metadata into a file that then leaves the organisation.
The reverse direction matters as much: a contract or policy document received from outside is untrusted input. Its body text, comments, and field codes are all places an instruction can sit where a human reviewer will not see it, and 'summarise this agreement' is exactly the request that puts it in front of a model holding other tools.
As with the other document skills, the reviewable artifact is the bundled code plus the instructions, and the useful control is where it runs rather than whether it runs.
Capability profile
What this skill can reach inside your environment, and why a security team cares. Derived from the publisher's documentation, not from a scan.
- bundled script execution
Executes with the invoking user's permissions to manipulate document packages.
- hidden content access
Tracked changes, comments and document properties preserve deleted text and authorship that the visible document does not show.
- document generation
Produced files can carry metadata onward and are often treated as authoritative once they leave the building.
- untrusted document ingestion
Body text, comments and field codes are all viable places to hide an instruction from a human reviewer.
Before you distribute this to a team
- Strip metadata before external distribution
Agent-generated documents inherit whatever the pipeline leaves in them. Check properties and revision marks before anything goes outside.
- Treat received documents as untrusted input
Comments and tracked changes are unread by humans and fully read by models.
- Review and pin the bundled scripts
Read the code, pin the version, re-review on change.
Vincosha Assay assessment
not yet assessedWe have not published a scan verdict for Word (docx) skill.
Everything above is a capability profile compiled by hand from the publisher's own public documentation on 2026-07-25. It describes what this artifact can reach and what to restrict before rolling it out. It is not a security verdict, and it should not be read as one.
When Vincosha Assaypublishes an assessment for this artifact it will appear here with a verdict, the findings behind it, and the date it ran. We would rather leave this space empty than fill it with a guess — inventing a security verdict about somebody else's software would be both wrong and dangerous.
What an assessment covers
- prompt-injection patterns in descriptions and instructions
- credential reads and outbound egress in bundled code
- over-scoped permissions and manifest claims
- typosquatted and impersonating names
- tool-manifest drift since the last approval
- content hash against the version you approved
Sources
Compiled from the publisher's own documentation and repository. Skills change — read the version you are about to install rather than trusting a profile.