The below works for me. As I said above the 4th order equations change very rapidly in time, so it makes sense T=8 (seconds) might be enough to cause approximation errors (or very small vibrations, unmeasurable).
I'm was hoping after my "brief hint" (see above) that you'd see the inequality of y[x,0] and y'[0,t] and adjust them, you did with (Cos[x]-1). Great!
c = 5;
L = 2 Pi;
T = 1/100;
(*EQUATION*)
PDE = D[y[x, t], {t, 2}] + c^2*D[y[x, t], {x, 4}] == 0;
(*INITIAL CONDITIONS*)
IC1 = y[x, 0] == Cos[x] - 1;(*Initial shape of the beam*)
IC2 = D[y[x, t], {t, 1}] == 0 /. {t ->0};(*No initial velocity*)
(*BOUNDARY CONDITIONS*)
(*Clamped on the left side of the beam*)
BCLeft1 = y[x, t] == 0 /. {x -> 0};
BCLeft2 = D[y[x, t], {x, 1}] == 0 /. {x -> 0};
(*Clamped on the right side of the beam*)
BCRight1 = y[x, t] == 0 /. {x -> L};
BCRight2 = D[y[x, t], {x, 1}] == 0 /. {x -> L};
(*SOLVE*)
soln =
NDSolve[{PDE, IC1, IC2, BCLeft1, BCLeft2, BCRight1, BCRight2},
y[x, t], {x, 0, L}, {t, 0, T}][[1, 1, 2]]
Using MaxStepSize, which the Mathematica NDSolve Message suggested, I was able to plot up to T==2. I was unable to do T==8 using MaxStepSize.
NDSolve[{PDE, IC1, IC2, BCLeft1, BCLeft2, BCRight1, BCRight2},
y[x, t], {x, 0, L}, {t, 0, 2}, MaxStepSize -> .01][[1, 1, 2]]
If instead of MaxStepSize I set c = 5/(10^2) then T==8 is allowed, I think because the solutions on the broken plot are10^32 (and growing).
"lastly" I tried MaxStepFraction -> 1/100, and that worked like your second choice did (the softer lobe plot, Method -> {"Adams", "MaxDifferenceOrder" -> 1}).
Why? As you know the most basic ODE approx. method has the problem: the further time gets from the initial values provided the worst the error is. For PDE boundary problems the discretization of the boundary effects error and maybe time evolutions does too. I am not familiar with all the methods, nor have I hand approximated this problem, to say more.