The recent suggestion that earthquakes it Long Valley caldera, California are caused by rapid tensile failure under high fluid pressure contradicts the assumption that the speed of propagation of a fluid-driven tensile crack is limited by the speed of motion of the fluid with the crack, so that such cracks cannot radiate seismic waves. In this paper we analyze fluid-driven cracks of various geometrics and find examples of both stable and catastrophic propagation. For an isolated crack in a homogeneous medium with remotely applied compressive stresses the stress intensity decreases as the crack tip extends beyond the driving fluid (assumed stationary), so that crack propagation is stable and limited by the rate of fluid flow. On the other hand, for two cracks approaching each other, initially stable behavior gives way to unstable propagation, and the cracks join catastrophically even if the fluid is stationary. A crack approaching a free surface also exhibits a transition from stable to unstable propagation, with the final episode occurring catastrophically. A crack emanating from a pressurized reservoir of cylindrical or spherical shape can also undergo an episode of catastrophic propagation. Propagation begins when the reservoir pressure rises to a critical value, which depends on the regional stress field, the fracture toughness of the rock, and the size of the largest flaw at the surface of the reservoir. This flaw propagates catastrophically until it reaches a certain aize and thereafter propagates stably. Microearthquakes caused by hydraulic fracturing have dimensions of at most a few tens of borehole radii and therefore probably are difficult to detect. Volcanic earthquakes can have source dimensions comparable to the radius of the associated magma chamber, so earthquakes as large as those at Long Valley caldera (source dimensions ≂ 10 km) could be caused by tensile failure. ¿ American Geophysical Union 1987