The source establishes a concrete automatic agent-extension path from workspace control messages to local skill installation. It does not establish covert exfiltration, install-time mutation, or a completed malware chain.
Static reason
High-risk behavior combination matched malicious policy.; previous stored version diff introduced dangerous source
Trigger
A connected workspace sends a skill installation control action.
Impact
A party able to issue workspace control actions can add instructions or capability-bearing skill content to a configured agent, creating a guarded agent-extension risk.
Mechanism
The handler downloads supplied skill content or fetches a GitHub repository, then writes it into an agent skills directory; the Claude adapter can launch with permissions skipped.
Evidence
src/adapters/base.jssrc/skill-installer.jssrc/adapters/claude.js
Network endpoints2
github.com/raw.githubusercontent.com/