They refer to either an attack in which (in optimal conditions) the channel is closed you cannot retrieve the funds because your "retrieve your funds out of attackers hands" transaction comes to late in the blockchain.

If you know how #lightning works, you know that the #Alice could potentially broadcast an earlier state of a transaction channel to settle funds she actually paid to #Bob in the most recent channel state, but because #Bob has the proof that #alice paid him, he can broadcast that so the others reject alices attempt.

He has a certain period to do that, because of a cryptographical setup that makes it that the attacker always needs to wait longer then the victim to retrieve funds. But if #Alice is able to stop #Bob from broadcasting that "save my funds" transaction, it could be that Alice is able to make her broadcast first.

She cannot stop #Bob from making that transaction, but she can make some kind of fake transaction that looks like bob's one and thus replace it in the memory pool of other nodes.

If she succeeds, she is able to broadcast her retrieving funds transaction before #Bob can (his one gets ignored because a fake replica takes its place in.) and thus steal the funds.

This attack is really sophisticated and cannot be mass utilized. It also requires optimal conditions and there are counter measures against it.

it is called something like "lightning replacement cycling attack".

The other case is that they refer to a recent issue with LND in which a node can get spammed and therefore forceclose channels in which case you need to retrieve your funds on-chain. But this isn't nice because you need to pay a transaction fee which can be high.