How to Govern OpenAI Assistants
The Assistants API handles tool calls and code interpretation. Cordum adds the governance layer that OpenAI does not provide.
The problem with ungoverned OpenAI Assistants agents
- Assistants execute function calls based on model reasoning. You define functions but the model decides when to call them.
- Code Interpreter runs arbitrary code without organizational policy enforcement.
- The API gives you runs and steps but no structured policy decision log.
- When an assistant calls the wrong function, the only safety net is the function implementation.
How Cordum governs OpenAI Assistants
Proxy assistant runs through Cordum
Submit OpenAI Assistant tasks as Cordum jobs. Safety Kernel evaluates the function set and user context.
const job = {
type: "openai.assistant.run",
payload: {
assistantId: "asst_abc123",
functions: ["search_database", "send_notification", "update_record"],
userMessage: userInput,
},
metadata: { user: userId, department: "support" },
};
await cordumClient.jobs.submit(job);Restrict function access by context
Policies control which functions an assistant can call based on user role, department, or time constraints.
Gate sensitive function calls
REQUIRE_APPROVAL for functions with real-world side effects. Execution pauses until approved.
Redact or quarantine outputs
Output safety layer can REDACT sensitive data or QUARANTINE responses matching risk patterns.
OpenAI Assistants native vs Cordum governance
| Area | OpenAI Assistants Native | With Cordum |
|---|---|---|
| Pre-execution function gating | Model decides function calls | Policy check before function execution |
| Context-aware access control | Same functions for all users | Per-user, per-role function access policies |
| Approval workflows | requires_action status (manual) | Declarative approvals with approver routing |
| Output safety | None at function level | ALLOW / REDACT / QUARANTINE on outputs |
| Policy versioning | N/A | Version-controlled policy bundles with snapshots |
FAQ
Does Cordum replace the Assistants API?
No. Cordum wraps the assistant lifecycle with governance. Your assistant still runs on OpenAI.
Can I govern Code Interpreter?
Yes. Policies can match on code_interpreter and constrain or require approval.
Related guides
Ready to govern your OpenAI Assistants agents?
Start with the open-source Cordum platform. Add policies, approvals, and audit trails in minutes.