Security architecture

Speedwave connects autonomous AI agents to mission-critical developer tools, production repositories, and messaging platforms. Because language models can encounter malicious inputs or prompt injection payloads, Speedwave’s architecture is engineered around the principle of least privilege, ensuring that potential compromises remain strictly contained.

These foundational security invariants govern the entire platform:

Foundational security guarantees

• Zero credentials in the model sandbox: The container executing Claude Code holds zero external service tokens and has no access to the container management socket. If the agent process is compromised, no long-lived secrets can be exfiltrated.
• Strict credential segregation: Each service integration executes inside an isolated worker container mounting only its designated service credentials in read-only mode. A vulnerability in the Slack worker cannot compromise GitLab tokens.
• Fail-closed startup gates: The SecurityCheck validation engine inspects configuration manifests prior to container startup, blocking execution if secret patterns or container engine sockets are exposed.
• Dynamic data tokenization: High-entropy secrets and sensitive data returned by integrations are masked with deterministic cryptographic tokens before payloads reach the LLM provider.

Multi-layered defense-in-depth

Speedwave enforces multiple overlapping boundaries:

1. Sandboxed execution: Claude Code executes within an unprivileged container with dropped Linux capabilities and a read-only root filesystem.
2. Virtualization boundary: On macOS and Windows, containers execute inside an isolated Linux virtual machine (Lima or WSL2), separating the container network from the host operating system.
3. Stateless Tool Gateway: The routing mediator holds zero credentials, mitigating risk in the event of routing-layer inspection.
4. Binary verification: External binaries and container base layers are validated against cryptographic digests prior to execution.

Explore security domains

Isolation model Container boundaries, unprivileged execution, and network segregation.

SecurityCheck engine Fail-closed configuration validation and runtime security audits.

Data tokenization Deterministic AES-128-SIV pseudonymization of sensitive values.

Binary authenticity Apple Developer ID code signing and Ed25519 update verification.