The threat of a 51% attack still exists on proof-of-stake as it does on proof-of-work, but it's even riskier for the attackers. A attacker would need 51% of the staked ETH (about $15,000,000,000 USD). They could then use their own attestations to ensure their preferred fork was the one with the most accumulated attestations. The 'weight' of accumulated attestations is what consensus clients use to determine the correct chain, so this attacker would be able to make their fork the canonical one. However, a strength of proof-of-stake over proof-of-work is that the community has flexibility in mounting a counter-attack. For example, the honest validators could decide to keep building on the minority chain and ignore the attacker's fork while encouraging apps, exchanges, and pools to do the same. They could also decide to forcibly remove the attacker from the network and destroy their staked ether. These are strong economic defenses against a 51% attack.

Staking-as-a-service will operate as a market of diverse options. The network could decide to fork from an outsized dishonest market participant. It would be controversial, but the validity of the network is generally a common incentive for most involved.