If you bring in the gain-phase plot for the TL431, you can figure it out.
I would suggest one thing, for two reasons. An emitter resistor. First, this regulates the voltage gain of the transistor, so that rather than having transconductance proportional to current (and rather high to begin with, because BJTs are awesome), it's limited by that resistor. Second, this allows some voltage swing to appear at the TL431, so if nothing else, you can compensate it locally with a C or R+C back to REF. (Without the emitter resistor, the emitter voltage will be largely the same as the base voltage, less a fixed drop, so the feedback C or R+C will simply feed back no voltage.)
Note that this circuit doesn't have straightforward gain and phase and all that, because there's no independent collector load, it's all shunted back to the same point. Likewise, base voltage is set by collector voltage, so the transistor looks more like a fixed gain noninverting (not really "common base" anymore) stage.
On the upside, the limited gain will probably correct the ills of the transistor to a frequency much higher than the TL431 understands (i.e., the roll-off for the gain shown will be in the 10s of MHz), and one should be able to make the hand-waving assumption that, as long as feedback gain is less than transistor gain, it will be stable (because, dividing the two out, the TL431 is effectively still running at or above unity gain, and we know it is unity gain stable, so it should still be stable). That's no guarantee, but it's a start.
BTW, SPICE models of the TL431 are notoriously bad, even as simplified macro models go. Info:
http://www.audio-perfection.com/voltage-regulators/linear-regulators/realistic-spice-model-for-tl431-stability-noise-impedance-and-performance-simulation-of-tl431-shunt-regulator.htmlHelmut is very knowledgeable at LTSpice so, if that's the default model used, you're probably not too bad off. I would recommend that model, or the "Eugene" model given in the link.
Tim