Could Another Firm's Jailbreak Reach Your Data?

This is the question we get asked most often, so it is worth answering precisely. A jailbreak manipulates a model into saying something it should not, using the attacker's own conversation. It cannot reach across into yours, because there is nothing shared between the two to reach through.

Your prompts are not stored in the model, and the model's weights do not change when it answers you. There is no shared memory between requests, so there is no channel for one session to read another. This is a property of how inference works, not a policy we apply, which is why it holds no matter whose model you point your key at.

The risk worth planning for is different, and it sits inside your own engagement: a target system whose content carries instructions aimed at your agent. That one is real, and it is what our containment controls are built for.

The distinction that matters

• A jailbreak of a shared model reaches only the attacker's own conversation, never yours.
• Nothing you send is written into the model, so there is nothing there to extract later.
• A breach of an AI provider exposes what that provider holds. Because the account and the agreement are yours, that exposure and its notification duties sit with your provider, under terms you signed.
• If your contracts do not permit a third-party AI processor at all, you can run the models inside your own cloud account instead.
• Hostile content inside a target's scope is the live risk. See how injected instructions are contained.