Compute security and data handling
How to set up your data, credentials, regions and untrusted inputs before using Compute, with Trust holding the deeper evidence.
Decide what you control before transfer
Before sending customer work, choose the data profile, local preparation rules, credential arrangement, permitted placement, and destination. Lithi checks those choices before transfer and execution. They limit what Compute may process and where it may go, but they do not make a participating Mac a confidential-computing enclave or make prepared data anonymous.
Start with the guide that matches the decision you need to make:
- Data profiles — what kinds of data the selected processing profile permits or refuses.
- What leaves your machine — what local preparation checks and decides before anything is sent.
- How your job is split across Macs — the owner cap and spread-placement limits on a multi-piece job.
- Credentials — give each client only the access it needs and rotate/revoke it deliberately.
- Regions — apply supported region/jurisdiction restrictions.
- Untrusted inputs — keep document/tool content from becoming authority.
Understand the deeper security evidence
Compute Docs explain what to configure. The Trust Center owns the deeper evidence and limitations:
A participating Mac is not a confidential-computing enclave. Preparation, encryption, narrow permissions, processing profiles, and the runtime security model bound exposure. Sharding limits how much of a multi-piece job one owner can reach; it does not make the assigned plaintext unknowable to a privileged host owner. See how your job is split across Macs for the exact boundary.
For the design behind these choices, read Prepare and protect your input and Split work without losing meaning or privacy in Lithi: Crowdsourced AI compute, powered by people. The paper's limitations section explains what those boundaries do not prove. After an accepted run, use Read and verify Compute result files to check the delivered artifact and its evidence.
Use Help for a security symptom
If authentication fails or a credential may be compromised, use Account/API key help or Authentication help. Report a suspected vulnerability privately through Report a security issue.