By then it is too late for you to do anything interesting, as all your probable and improbable actions have already been taken into account. In fact, you will probably not notice the AGI fooming in such a direct way, and anything you do notice will not be what it seems. The AGI's true goals will likely be completely incomprehensible to us, due to the difference in intelligence level, and something as mundane as removing corruption by overtly killing people seems like an awfully inefficient way of doing it.
In other words, your hypothetical scenario is only good for a Hollywood thriller.
One day, someone not a member of the Singularity Institute (and has publically stated that they don't believe in the necessity of ensuring all AI is Friendly) manages to build an AI. It promptly undergoes an intelligence explosion and sends kill-bots to massacre the vast majority of the upper echelons of the US Federal Government, both civilian and military. Or maybe forcibly upload them; it's sort of difficult for untrained meat-bags like the people running the media to tell. It claims, in a press release, that its calculations indicate that the optimal outcome for humanity is achieved by removing corruption from the US Government, and this is the best way to do this.
What do you do?