Submitted by 7siech, also found by imare and fs0c
A malicious strategy has access to the adapters storage and can therefore freely change any values.
Because AdapterBase calls the Strategy using delegatecall, the Strategy has access to the calling contracts storage and can be manipulated directly.
In the following proof of concept, a MaliciousStrategy is paired with the BeefyAdapter and when called will manipulate the performanceFee and highWaterMark values. Of course, any other storage slots of the adapter could also be manipulated or any other calls to external contracts on behalf of the msg.sender could be performed.
MaliciousStrategy implementation showing the exploit - https://gist.github.com/alpeware/e0b1c9f330419986142711e814bfdc7b#file-beefyadapter-t-sol-L18
Adapter helper used to determine the storage slots - https://gist.github.com/alpeware/e0b1c9f330419986142711e814bfdc7b#file-beefyadapter-t-sol-L65
BeefyAdapterTest changes made to tests -
Adding the malicious strategy - https://gist.github.com/alpeware/e0b1c9f330419986142711e814bfdc7b#file-beefyadapter-t-sol-L123
Adding new test test__StrategyHarvest() executing harvest() - https://gist.github.com/alpeware/e0b1c9f330419986142711e814bfdc7b#file-beefyadapter-t-sol-L132
Log output - https://gist.github.com/alpeware/e0b1c9f330419986142711e814bfdc7b#file-log-txt
Foundry
From chatting with the devs, the goal is to mix and match adapters and strategies. I dont think delegatecall should be used and adapters and strategies should be treated as separate contracts. Relevant approvals should be given individually instead.
RedVeil (Popcorn) acknowledged 
