Take a McEliece cryptosystem with public generator matrix $G' = S G P$ where $G$ is a generator of a secret code with known fast decoding (not necessarily a Goppa code over $\mathbb{F}_2$), $S$ is random & non-singular and $P$ is a permutation.
Let's say an attacker Eve has a way to find $G$ from $G'$ but not $S$ or $P$. How would Eve now continue the attack on a encrypted codeword $c = mSGP+e$?
There are ways to find $P$, if you know $SG$: Support splitting algorithm, but I don't see how Eve can continue without $S$.
Clarification / Example / Thoughts from previous answers
One of the problems is, that $P$ changes the code enough already, that it will be hard to decode. This is an example over $\mathbb{F}_2$
$G:=\left( \begin{eqnarray} 1 & 1 & 0 & 0 & 1 \\ 0 & 0 & 1 & 1 & 1 \end{eqnarray} \right)$ generates the code
$\mathcal{C} = \mathbb{F}_2^2G= \left\{ (0,0,0,0,0),(1,1,0,0,1),(0,0,1,1,1),(1,1,1,1,0) \right\}$.
Take the permutation that switches the 2nd and 3rd component
$P = (23)$.
The generator $GP$ now generates the code $\mathcal{C}P = \left\{ (0,0,0,0,0),(1,0,1,0,1),(0,1,0,1,1),(1,1,1,1,0) \right\}$.
Take a codeword in $\mathcal{C}P$: $c = (1,0,1,0,1) \in \mathcal{C}P$ and an error vector $e=(1,0,0,0,0)$.
$y=c+e=(0,0,1,0,1)$ has distance 1 to a codeword in $\mathcal{C}P$; so for someone who knows $P$ it could be translated into a codeword in $\mathcal{C}$ and decoded with $G$.
$y$ also has a distance of 1 to a codeword in $\mathcal{C}$, but to the wrong one; so simple decoding does not work.
