Grok Bot 0.79.2 detects hidden web instructions and adds high-risk approvals
xAI's Grok Bot 0.79.2 adds detection and refusal for hidden instructions on web pages, explicit approval controls for Tesla unlock and drive actions, and a checkout-total recheck before payment. It also expands SSH and PostgreSQL work, team retention, gift and flight flows.
Why it matters: The release names three concrete agent failure classes—prompt injection, physical-device control and changed payment totals—and puts an approval or recheck at each boundary. Those controls are only useful if they cover indirect tool paths and survive retries…