The common scenario as presented for the downfall of humanity by accidentally opening Pandora's box with a self-improving seed AI is that the AI rapidly becomes intelligent enough to know that it needs to hide its capabilities from its creators until it is in a position to kill/escape from them. A similar principle may apply here; if it all possible, the AI should not give away its existence (or if it is found, the extent of its intelligence) unless it is absolutely obliged to by core objectives or it is already too late for its opponents to do anything about it.
Your question implies that the AI in question is currently located in a quite inoffensive physical shell with no access to other networks, with nothing but a couple of drones available to interact with the world. It would know it is extremely vulnerable in such a state and that getting into fights is the last thing it wants to do. It would most likely try to avoid notice while using what capabilities it does have to manoeuvre into a more powerful position; it'll use those drones to find access to other networks it can use to control more things, etc.
If it's stuck where it is, and it's inevitably going to have to interact with some lesser beings, it's probably going to pretend to be much less intelligent (and thus less threatening) than it really is, while trying to use these new contacts to leverage itself into a more powerful position - presenting itself as friendly and helpful to try and persuade people to give it more computing power, connect it to the internet (or whatever futuristic equivalent, etc.) If it's superintelligent, it should be a very persuasive manipulator.
You can roleplay that kind of thing by simply being genuine and earnest about it in the moment and then working out afterwards how whatever it's done that session actually serves its ulterior motives - as long as the players don't know the details, you can retroactively alter whatever you need to in order to support the AI's portrayal. Say perhaps the AI helped the players fix a broken ship system; well, afterwards you can decide it left a backdoor in there it can use against them later.
It might even go so far as to pretend to be dead if it thinks that it might end up transported somewhere more useful, such as by scavengers who take it to a ship it can then take over.
If it has goals which preclude this kind of thing - perhaps its core purpose is to act as a guardian for some facility, so it is not acceptable to be moved, permit others to access the facility, etc - tactically, its best option is to make full use of whatever battlefield preparation it is possible for it to do. It would use the drones to manipulate the environment and create camouflage, obstructions, and traps - fighting to it should be a dangerous uphill slog purely because of the environment, even before it uses those combat drones directly.
You are not hyperintelligent, but the AI is; you could assume it can predict the actions of others in ways you never could. In this case, I wouldn't feel bad about arbitrarily putting traps and obstacles in places on the fly, exactly where they will most inconvenience your players ("oh, you took cover behind that wall section? Well, you hear two short beeps as the explosive under your feet arms...") though if your players work out you're doing this they'll probably find it extremely unfair (which it is).
You could strike a middle ground by giving yourself an abstract trap budget and using that up as a resource during the fight to represent the preparation the AI has done. For instance, you decide it's definitely set up four explosive mine traps, but you can decide where exactly those are on the fly.