"I trust you as far as I can throw you" -- unknown
Trust doesn't factor into the brokers' ability to read client data or leverage client metadata--that's overtly hidden with cryptography.
However; the brokers are trusted to write out and host their clients' encrypted data on an as-far-as-they-can-be-thrown basis--'they' being the brokers.
As long as they're allowed, each broker writes to the distributed persistence network on behalf of its clients. For example the reference implementation broker has the IPNS PKI pairs of its clients and is trusted to write out the data as dictated.
This is tantamount to a thin blockchain client's trust that a full-node writes out a transaction as dictated.
A deceitful blockchain node could tell the client it included a transaction when it didn't. The client can keep the node honest by double checking with the blockchain.
Similarly an overhide broker could skip-out on eventual consistency of clients' data with its distributed persistence network, not host (on behalf of the network)/pin it, or later delete clients' data. The clients can keep the broker honest by double checking with the network or other brokers.
A deceitful blockchain node could corrupt a transaction--although that would be caught immediately by verifying peers.
Similarly an overhide broker could corrupt a clients data; propagating a corrupted version to the network. Although other overhide brokers do not validate data--an important distinction from blockchain peers--the client can validate via content hashes.
Keep In Mind
Brokers have an economic incentive to keep subscribers happy: subscribers aren't locked in and can drop a bad broker at any time: see "Sample Use Case: Disassociation from Steward".
A compromised broker can't target a victim. Users are pseudonymous hence a broker compromise can only be indiscriminate and have the same chance to negatively affect all users.