Can anomalies exist without gauge fields?
I would defer to the extant answer, only to adapt its item 2 to your potentially narrower question. Matt S is somewhat glib here, in that there would be something in theories without [manifest, overt] gauge fields. In low energy effective hadron theories with anomalous chiral s.broken symmetry, the (4d) WZW sigma model term does the trick of providing the r.h.side to the divergence of the anomalous current.
If a Martian landed here and talked to a low energy data-analyzer without the benefit of a theorist (who'd know about anomaly consistency conditions and the like), and were curious about couplings such as KKπππ inaccessible through the standard chiral model, he might be then be served the WZW term of the effective theory that provides it. Where? he might ask.
The earthling would point to the leading term of the WZW term of low energy physics, silent about how it was produced, lacking manifest gauge fields, $$ \frac{2N_c}{15\pi^2f^5}\int_{S^4} d^4x ~ \epsilon^{\mu\nu\kappa\lambda} \operatorname {Tr} (\pi \partial_\mu \pi \partial_\nu \pi \partial_\kappa \pi \partial_\lambda \pi), $$ the πs being are SU(3) flavor octet axial-goldstons, $\pi\equiv \pi^i \lambda^i/2$ with their adjoint eightfold-way indices compacted into flavor Gell-Mann matrices.
These 8 goldstons shift under the 8 sbroken axial transformations, $$ \delta_{\theta_A } \pi\propto \theta_A ~ v_{\chi SB} + ... $$ with higher order goldston pieces.
He'd immediately see, to leading order in goldston fields, axial invariance for a hyperspherical closed boundary, since the only surviving variation of the lagrangian would be the constant underived term inside the trace, so the lagrangian would be a surface term, vanishing on the closed hyperspherical boundary.
So... what is the axial current octet corresponding to these shifts? He'd compute an extra piece beyond the standard linear one, $$ \theta_A\cdot J_{A\mu}=\frac{-8N_c}{15\pi^2f^5} \epsilon^{\mu\nu\kappa\lambda} \operatorname {Tr}(\theta_A ~\pi \partial_\nu \pi \partial_\kappa \pi \partial_\lambda \pi ), $$ diverging to $$ \partial^\mu (\theta_A\cdot J_{A\mu})=\frac{-8N_c}{15\pi^2f^5} \epsilon^{\mu\nu\kappa\lambda} \operatorname {Tr}(\theta_A ~\partial_\mu\pi \partial_\nu \pi \partial_\kappa \pi \partial_\lambda \pi ). $$ The r.h.side is non-vanishing, without manifest gauge fields, and (behold!) a total divergence, so, then, a surface term. Without manifest extraneous gauge fields. It leads with a quadrilinear term, but of course higher orders cannot help comporting with the pattern. That is, as the answer above indicates, the anomaly is a feature of the functional integral transformation structure and not of the extraneous handles that latch on to the relevant currents.
There are three types of anomalies:
't Hooft anomaly: This is a potential anomaly. The symmetry is exact, global, but connot be consistently gauged. This type of anomaly occurs for example in the flavor symmetry $SU(N_f)_L \times SU(N_f)_R \times U(1)_V$. This anomaly is important because it effects the low energy content of the theory due to the 't Hooft anomaly matching condition which states that the anomaly should be the same when computed at high or low energies.
The chiral anomaly. In this case a vector symmetry is made local and the corresponding current couples to a quantized gauge field. In this case due to $VVA$ triangle diagram, there is an anomaly in the axial current, which will not be conserved, but since the axial current is not coupled itself to a gauge field, it does not need to be conserved and the theory is consistent.
The chiral gauge anomaly. In this case an axial symmetry is gauged and the corresponding axial curent is coupled to a quantized gauge field. In this case the $AAA$ triangle diagram results an anomaly in the axial current and it will not be conserved; but it is gauged thus must be conserved in order to couple to a gauge field. This is the dangerous case where the theory becomes inconsistent. This is the case in the standard model where the axial current is coupled to the vector bosons. Here the anomaly must be canceled for the theory to be consistent. The cancelation can be taken care of by an educated selection of the fermion species representations as in the standard model, or the cancelation of the bulk anomaly with the edge anomaly as in the quantum Hall effect.