Airlock Supports Bring Your Own Fireworks AI Key

Airlock has announced a new capability: users can use their own Fireworks AI API key to power Airlock's governance action review. This feature is currently in early access.

What Is "Bring Your Own Key" Mode

"Bring Your Own Key" (BYOK) means that instead of relying on the platform's built-in AI usage quota, users initiate model requests with their own Fireworks AI key. For Airlock, this means the large model calls behind review actions are carried by the user's own Fireworks AI account.

This model is becoming increasingly common in AI governance products. It decouples model usage costs from platform subscriptions and gives enterprises more control over data flows, billing, and vendor selection.

BYOK (Bring Your Own Key) has a long history in cloud services and SaaS, initially used mostly for encryption key management scenarios and later widely adopted for AI model calls. In its typical architecture, the platform does not hold the user's vendor credentials in plaintext, but instead initiates requests on the user's behalf via short-lived tokens or encrypted storage. Users can revoke or rotate keys on the vendor side at any time, retaining ultimate control over credential lifecycle management. For auditing, BYOK lets call logs be verified directly under the vendor account, making it easier for enterprises to incorporate AI calls into their existing cost centers and compliance systems.

Governance Logic Stays on Airlock's Side

It's important to distinguish that bring-your-own-key only changes the source of model calls. Airlock explicitly states that policies, fixed rules, and human approvals remain managed within Airlock.

In other words, Fireworks AI provides the inference capability, while governance decisions—"which actions require review, what rules are triggered, whether human intervention is needed"—remain the responsibility of Airlock's policy engine and approval workflows. This division keeps governance boundaries clear: the model handles understanding and judgment, while the platform handles rules and accountability.

In the context of AI governance, a "policy engine" typically refers to a component that formalizes organizational rules (such as lists of sensitive operations, risk tiers, and trigger thresholds) and enforces them automatically. It is a separate layer from the large model responsible for semantic understanding. Policy engines emphasize determinism and auditability—the same input should produce a reproducible decision—while model inference is probabilistic and better suited to judgment tasks such as intent recognition and risk classification. Separating the two allows rule changes to go through approval and version management processes, while model upgrades or vendor switches do not affect the semantics of the rules themselves.

Who This Capability Is For

For teams already using Fireworks AI, bring-your-own-key avoids duplicate purchases of model quotas and makes it easier to unify billing and usage monitoring. For enterprises with data compliance requirements, using their own key means the account entity in the call chain is more clearly defined.

Early access means the feature may still be iterating. Before integrating, it's advisable to confirm the scope of key permissions, call quotas, and how Airlock specifically bills for review actions.

Fireworks AI is a platform that provides model inference services, focused on low-latency model hosting and inference APIs. It supports a variety of open-source and fine-tuned models and is often used in high-throughput production scenarios. For teams with an existing Fireworks AI account, billing is typically settled by token usage, and call details and quotas can be viewed in the console. This means that after integrating Airlock, the inference consumption generated by review actions will appear in the existing Fireworks AI bill rather than being added on top of the Airlock subscription.

FAQ

After bringing my own key, do Airlock's rules still need to be configured? Yes. Policies, fixed rules, and human approvals are all managed in Airlock; the key only affects model calls.

Is Fireworks AI mandatory? Based on the current announcement, this is an integration option for Fireworks AI keys. Support for other model vendors is subject to Airlock's official documentation.

What are the limitations of early access? The announcement did not disclose specific limitations. It's advisable to confirm quotas and support scope with Airlock before integrating.