Agentic Security Operations Orchestrator

Problem

The planned prototype would examine how probabilistic proposals could assist a security workflow without giving a model unrestricted authority to change external systems.

Target users

It would be intended for engineers studying safe agent orchestration and security-operations workflow design.

Why it matters

If implemented, it could make approvals, permissions, retries, duplicate prevention, and failure recovery explicit around any suggested action.

Main features

• Would model each workflow as explicit durable states
• Would expose narrow typed tools rather than generic system access
• Would require policy validation before execution
• Would pause higher-risk actions for human approval
• Would record idempotency, verification, and reconciliation events

System architecture

The planned design would keep the model as an untrusted planner while deterministic application code would own policy, state transitions, tool execution, and audit data.

Data flow

A simulated alert would enter a durable workflow; a model could propose a typed action, policy code would validate it, and an approved adapter could run only in a sandbox before verification.

Backend architecture

A future orchestrator would store state revisions, approvals, action intents, receipts, and reconciliation status.

Frontend architecture

A future operator view would show proposed actions, evidence, approvals, execution receipts, and recoverable failures.

AI and ML techniques

• Planned constrained action proposal
• Planned tool selection evaluation
• Planned adversarial prompt testing
• Planned human approval workflow

Deployment approach

The prototype would run against simulated systems with test credentials in an isolated environment, keeping all action paths sandboxed.

Security and privacy

The planned design would use least-privilege test credentials, tenant-scoped policy checks, destination allowlists, redaction, bounded outputs, and audited human approvals.

Evaluation strategy

The planned evaluation would inject malformed proposals, prompt injection, stale state, duplicate delivery, timeouts, denied permissions, and unknown external outcomes.

Engineering challenges

The project would need to distinguish transient failure from unknown outcome while preventing duplicate or unauthorized effects across retries.

Trade-offs

More approvals and narrower tools would reduce autonomy but could make the simulated workflow safer, more testable, and easier to audit.

Impact

If built, the prototype could demonstrate control boundaries for agentic security workflows through policy checks, approvals, bounded tools, and recoverable state transitions.

Future improvements

Later work could expand the simulated adapter set, add policy-version comparisons, test multi-step recovery, and study operator review quality.