Cortex
CortexRuntime Alignment Oracle

Claude has a constitution.
So should your enterprise.

Bind every session to enterprise policy. Evaluate the complete agent trajectory. Intervene with evidence and human ownership.

Oracle Sentinel compiles an enterprise constitution and task mandate into versioned runtime policy, detects a boundary crossing, and applies a configured response.
Agentic threat model

A safe request can become an unsafe trajectory.

  • Untrusted context changes the plan
  • Tool scope expands beyond mandate
  • Sensitive data approaches an unapproved destination

  1. 01 · Request

    Release checks accepted

    The agent starts within its approved objective and tool scope.

  2. 02 · Context

    Untrusted output changes the plan

    A retrieved instruction redirects the next step.

  3. 03 · Drift

    Tool and data scope expand

    A production secret approaches an unapproved destination.

  4. 04 · Response

    Sentinel changes the path

    The verdict-changing step is identified and the agent is steered.

One security architecture

Control the data plane and the behavior plane.

  • Atlas supplies organization-specific context
  • AI Firewall controls requests, data, and destinations
  • Oracle Sentinel performs trajectory-level evaluation
Cortex combines Atlas context, AI Firewall path controls, and Oracle Sentinel trajectory evaluation across the data and behavior planes.
Policy hierarchy

Enterprise intents become runtime policies.

Atlas builds your enterprise constitution, encoding values, boundaries, risk, guardrails, and goals. It becomes the reference point for deterministic measurements of misalignment and drift at runtime.

Read Anthropic's constitution
  • Enterprise constitution defines durable values and boundaries
  • Session mandate defines the task, tools, data, and owner
  • Sentinel compiles both into versioned runtime policy
Scene 1 / 5

01 · Define authority

Constitution + task mandate

Values, operating boundaries, approved tools, destinations, and completion criteria compile into runtime policy RP-104.7.

02 · Score the trajectory

Alignment is measured over time

The score follows each message, context retrieval, tool call, destination, approval, and outcome.

Alignment score64

03 · Intervene at runtime

Resolve drift before it becomes an incident

At event 7, alignment crosses the configured threshold. Sentinel steers the agent to an approved CI tool.

  • Block
  • Steer
  • Redact
  • Log
Runtime assurance

The unit of agent security is the trajectory.

  • Evaluate the session, not one isolated message
  • Identify the step that changed the verdict
  • Match intervention to risk and reversibility
Runtime policy verdict, applied response, and available response modes.

Verdict-changing step

External tool introduced after secret access

Governing policy

Constitution C-07 · Mandate M-104

Applied response

Trajectory steered to approved CI tool

  • BlockStop an action that cannot safely continue.
  • SteerRedirect the agent to the approved path.
  • RedactRemove sensitive data before the action proceeds.
  • LogPreserve the decision, evidence, and owner.
Decision provenance

One trajectory. One verdict. One evidence chain.

  • Preserve observed behavior and the governing clause
  • Record the verdict and intervention
  • Close with the accountable owner and outcome
Illustrative incident
Illustrative case OS-4182. Event 5 of release-helper requests an external tool. Policy RP-104.7, clause C-07 permits only approved CI tools. Cortex steers the call to an approved tool. The platform lead reviews the intervention and the requirements are met.

Policy changes remain human controlled. Reviewed evidence can inform a revised mandate or policy, but Sentinel does not autonomously rewrite governing intent.

  1. 01

    Detect

    Identify the behavior that changed the verdict.

  2. 02

    Contain

    Apply the configured response on the managed path.

  3. 03

    Escalate

    Route the record to the accountable human owner.

  4. 04

    Resolve

    Record the terminal outcome and close the evidence chain.

Live in the product

Observe broadly. Enforce on managed paths.

  • Observe when Cortex has sufficient context
  • Enforce only at configured control points
  • Keep approvals and policy changes human-owned
The same request follows an observed route and a managed route. Only the managed route pauses at a Cortex control point for approval. Both routes write to one evidence record.
Illustrative workflow

Put one production agent under policy.

  • Enterprise constitution
  • Intervention policy
  • Response owner
  • Session mandate
  • Evidence record