A BN-curve over a 256-bit prime field $\mathbb{F}_p$ has, being an elliptic curve, a 256-bit group attached to it, say of order $N$. As the best known attacks take $\approx\sqrt{N}$ times, this gives us 128-bits security against discrete logarithm attacks.
The curves also have embedding degree 12. That means we can use a pairing to map a discrete logarithm problem to $\mathbb{F}_{p^{12}}$. Given that $p\approx 2^{256}$, we know that $p^{12}\approx 2^{3072}$. Hence $\mathbb{F}_{p^{12}}$ is a 3072-bit finite field, and solving a DLP there should (see warning below) take about $2^{128}$ effort. So mapping the DLP to a finite field does not help an attacker, which is great.
If we move to 256-bit security, we will need a 512-bit group on the curve. Thus, we need to choose $p\approx 2^{512}$. The embedding degree remains 12, and we can still use the same pairing to map the DLP to $\mathbb{F}_{p^{12}}$. In this case $p^{12}\approx 2^{6144}$. According to the slides you refer to, this only gives a $<192$-bit security level. Thus $p\approx 2^{512}$ is too small for a 256-bit security level.
To obtain a 256-bit security level in the finite field, we actually need $p^{12}\approx 2^{15424}$ (according to the slides you refer to). Thus, $p\approx 1285$. This is very large, and will make computation very slow. Therefore BN-curves are great at 128-bit security level, as they make an optimal trade-off there, but do worse as the security level increases.
Warning: new attacks show that finite fields of size $2^{3072}$ do not actually provide a $128$-bit security level any more, although they did at the time of the proposal of BN-curves. So currently they are on shaky territory.