Security concepts

Alquimia Platform is designed for environments where AI agents handle sensitive data and perform real actions. Security is not a single feature; it is a set of layered controls that protect inference, data, and infrastructure.

Security layers

Identity & access

Authenticate callers and agents with API tokens, JWT, OIDC, and SPIFFE workload identity. Authorize actions by role and agentspace.

Secrets management

Resolve API keys, database credentials, and signing keys from HashiCorp Vault with dynamic, short-lived leases where possible.

Input & output safety

Shield classifiers inspect prompts and responses for injection, unsafe content, or policy violations before the main LLM is invoked.

Workload hardening

Run signed container images, enforce pod security, and verify supply-chain integrity with admission policies.

Agentspaces and isolation

An agentspace is a boundary that contains agents, secrets, and knowledge sources. It lets teams share a single Alquimia Platform deployment while keeping their configurations and credentials isolated. Each agentspace can have its own Vault policy, identity roles, and resource limits.

Zero Trust principles

Never trust, always verify — every request is authenticated and every secret fetch is attributed to an identity.
Least privilege — agents and runtime services receive only the permissions they need.
Assume breach — audit logs, metrics, and admission controls limit blast radius and detect anomalies.
Workload identity — services authenticate with short-lived credentials tied to their deployment, not static tokens.

Shields

Shields are guard models that run in the preprocess stage. They can:

Observe — record a verdict without changing the request.
Flag — warn about a potential issue while allowing inference to continue.
Block — stop the request before the main LLM is called.

A prompt-injection detector is typically configured in block mode so that a classification failure defaults to safe behavior.

Secrets lifecycle

Secrets are referenced by name in agent specs and resolved at runtime. This separation lets you:

Rotate credentials without redeploying agents.
Audit which agents accessed which secrets.
Use dynamic credentials for databases and caches.
Scope secrets per agentspace or share them globally.

Designing a security posture

When planning an Alquimia Platform deployment, decide:

Which authentication methods your callers and services will use.
How agentspaces map to teams, environments, or trust boundaries.
Which secrets should be dynamic and which can be static.
Which shields are required for each agent and channel.
How workloads will be hardened and images verified.
Who will respond to security alerts and how.