Module · Agent Security Framework

    Agents you can defend in a hearing.

    A security envelope built for regulated agentic AI: cryptographic identity, brokered secrets, controlled egress, prompt-injection defence, runtime detection, and a signed intent-to-action chain that a regulator can walk end-to-end.

    You choose the AI. Cabier orchestrates the trust.

    Request a private briefAgent WorkforceSovereign Intelligence Fabric

    Six domains

    Identity. Secrets. Egress. Context. Injection. Detection.

    The framework maps to NIST AI RMF, ISO 42001, MITRE ATLAS, and the EU AI Act Article 15 cybersecurity duty — and produces the evidence a supervisor actually asks for.

    Agent identity

    Every agent has a cryptographic identity, a scoped role, and a rotation policy. No shared service accounts, no ambient credentials.

    Secret & key hygiene

    Model API keys, tool credentials, and evidence tokens are brokered per-invocation. Nothing sits in a prompt, a log, or a client repo.

    Network egress

    Agents run inside a client tenancy or Cabier enclave. Egress is allow-listed per tool and per jurisdiction — enforced by the Sovereign Intelligence Fabric.

    Prompt & context isolation

    Prompts, retrieved evidence, and tool output are memory-isolated per tenant. No cross-client bleed, no vendor training reuse.

    Prompt-injection defence

    Untrusted content is quarantined and stripped of instruction semantics before it reaches a reasoning step. Tool calls require signed intents.

    Runtime detection

    Every invocation is monitored for jailbreak, exfil, and off-scope tool use. Anomalies trip a Trust Gate before an action lands.

    Control envelope

    What every agent runs inside.

    Signed intent, signed action

    A Tier 1 orchestrator signs the intent; the agent signs the action; a Tier 0 executive signs the release. Three signatures, one audit ID.

    Model-agnostic hardening

    The security envelope is identical whether the client runs GPT-5, Claude 4.5, Gemini 3, Mistral, or a sovereign small model. Swap the model — the controls do not move.

    Kill switch, per domain

    Every agent tier and every tool integration has a per-domain and global kill switch owned by a named executive. Trust Gates keep blocking even when drafting agents are paused.

    Regulator-ready evidence

    Identity, secret, egress, prompt, and detection logs are pulled into the same evidence vault that feeds ORS, CCM, and the Regulator Digital Twin.

    Threat model

    What the framework blocks.

    Prompt injection & indirect injection
    Untrusted retrieved content is treated as data, never as instructions. Tool calls require a signed intent that the injected text cannot forge.
    Model / weight exfiltration
    Model access is brokered through the CABIER AI control plane. Weights, embeddings, and fine-tunes never leave the client's residency envelope.
    Data poisoning & retrieval poisoning
    Evidence sources are signed at ingest by the Sovereign Intelligence Fabric. A poisoned or unlabelled packet cannot reach a reasoning step.
    Tool abuse & lateral movement
    Tools are allow-listed per agent, per scope, per jurisdiction. Off-scope tool calls trip a Trust Gate and freeze the workflow.
    Excessive agency & silent autonomy
    No agent can sign, publish, or transact. Autonomy is granted per task and expires with the audit ID.
    Insider misuse of an agent
    Every human action against an agent is attributed to a named identity, MFA-bound, and reviewable by a Tier 0 signer.

    Guardrails

    What the framework will not do.

    No shared agent identities
    Every agent has a unique, rotating, cryptographic identity. Shared service accounts are refused at deploy time.
    No prompts in logs
    Prompts, retrieved evidence, and tool output are redacted from operational telemetry. Full context lives only inside the signed audit ID.
    No unbrokered model calls
    All model traffic flows through the CABIER AI control plane. Direct client-to-vendor calls are blocked by policy and by network.
    No silent updates
    Model, prompt, tool, and policy changes are versioned. A Tier 0 signer approves a new version before it can serve a regulated workflow.
    Request a private briefAI Governance OSTrust Gates