Gemini Agent Hooks: Deny Risky Tool Calls in Python

On September 3rd OpenAI shipped GPT-6 Astra and, in the same breath, said it had crossed the Critical threshold for cybersecurity under its own Preparedness Framework. A model that scores 100% on ExploitBench is a model that can do things. The week's loudest developer conversation has not been about which model wins a benchmark. It has been about the boring question underneath: when your agent decides to run a shell command inside a sandbox, what stands between that decision and the filesystem?

Google's answer landed quietly on the same news cycle. Gemini 3.8 Flash went GA on September 2nd and became the default model behind the Antigravity managed agent, and the Hooks documentation was refreshed on September 4th. Hooks let you run your own script inside the agent's remote sandbox, immediately before or after a built-in tool executes, and return a one-word verdict: allow or deny.