Grok Build fixes permission, tool-safety bugs
The latest Grok Build update makes hook permission prompts appear regardless of approval mode and aligns WebSearch, WebFetch, image generation, and “Allow once” with configured policies and rule decisions. Grok Build is xAI’s terminal-native coding agent for complex development workflows.
This is unglamorous but essential agent infrastructure work: permission systems are only trustworthy when every tool follows the same authorization path.
- –Hook enforcement now remains visible across approval modes, reducing the risk of silent policy bypasses.
- –WebSearch, WebFetch, and image generation receiving consistent rule handling matters because external access and media tools can expand an agent’s attack surface.
- –“Allow once” is the right default for ambiguous or high-impact actions, especially in repositories containing secrets or deployment credentials.
- –The update reinforces Grok Build’s move toward a serious coding-agent platform with hooks, MCP, sandboxing, and configurable permissions.
DISCOVERED
1h ago
2026-10-02
PUBLISHED
1h ago
2026-10-02
RELEVANCE
AUTHOR
XFreeze