A security task earns its value when a team can hand a suspicious finding to a maintainer and provide reproducible evidence for a fix. A model saying it found a vulnerability is only a starting point. Daybreak and Claude CVP can reduce friction for legitimate defensive work, but neither determines asset ownership, deployment authority or final acceptance criteria for your team.
This guide separates product functionality, infrastructure requirements introduced with a new release, and our proposed operating method. Official facts were checked on 2026-11-07. API request bodies and prompts are unexecuted examples. The Python example handles fictional identities in memory: it has no network, repository edits, real customer information or attack payloads.
1. Start with capabilities: what defensive work do these programs support?
Daybreak is OpenAI’s cybersecurity access framework: Blue supports routine authorized defense, while Red provides separately approved specialist capability. Claude CVP organizes access into Defense, Red Team and Specialized tiers. Both aim to make eligible security work easier to perform with models. They are separate programs with separate controls, rather than interchangeable verification badges.
ChatGPT Learn: Models and Trusted Access · Claude Help: Cyber Verification Program
| Defensive task | Useful deliverable | Boundary to retain |
|---|---|---|
| Code review and threat modeling | Candidate findings tied to files and evidence; trust-boundary description | Static suspicion does not establish exploitability or exposure |
| Detection and incident analysis | Redacted timeline, detection-rule draft and unresolved hypotheses | Use approved logs without connecting unrelated systems |
| Remediation and validation | Minimal patch proposal, regression criteria and residual risks | People approve changes; passing tests does not establish system-wide safety |
| Customer security services | Reviewed report or patch deliverable | Check vendor program approval as well as the customer’s authorization |
A maintainer who already has a faulty function may need evidence organization and regression criteria rather than the most privileged model. Start with low-risk work through an already authorized surface. Consider a higher tier when the task needs it and the team can meet its controls. The workflow below is an engineering proposal, not a vendor recommendation to buy a particular subscription.
2. What changed on 5 November, and what did not?
OpenAI’s public API changelog dates the GPT-6-Cyber release to 5 November. The current model page describes vulnerability investigation, reproduction in an authorized test environment and fix verification; its model ID is gpt-6-cyber. This release date does not establish access for your ChatGPT, Codex, Work or API project.
OpenAI API changelog: GPT-6-Cyber on 5 November · OpenAI: GPT-6-Cyber model
Compared with the readable official documentation saved on this site on 4 November, the current Daybreak guide adds SSO/MFA requirements for Red API organizations and a dedicated GPT-6-Cyber setup section. Its table lists the new model as requiring Red plus additional model approval. The generic Models and Trusted Access body is unchanged. These captures establish a documentation difference within the interval, not the first effective timestamp of every rule.
OpenAI: Daybreak Responses API guide · ChatGPT Learn: Models and Trusted Access
A model upgrade can change capability, cost, tool combinations and operating prerequisites together. Replacing a model string in an old request covers only one migration step. A stronger comparison pins the model, access program, project, policy and input version for a task, then evaluates the same acceptance conditions. Record untested differences as unresolved; fewer refusals do not demonstrate greater accuracy.
3. New Red projects: carry approval into real infrastructure
An API project using GPT-6-Cyber needs Daybreak Red, additional model approval and project enablement. The official guide also requires an IP allowlist for API egress. A ZDR project needs PSP, storage marked Validated, and the Zero Data Retention with Private Safety Processing project policy. All Red API organizations must require SSO and enforce MFA through their identity provider.
| Control location | What to establish | What it does not establish |
|---|---|---|
| Organization and project | Approved model, enabled project and project-bound credentials | Browser identity verification does not prove API organization approval |
| API entry point | Public source IP reaching the provider matches the active allowlist | An API allowlist does not restrict an agent’s outbound destinations |
| Execution environment | Read-only repository copy, no network and controlled tools | An instruction to stay in scope does not replace permissions |
| Data controls | Exact project retention policy, validated storage and accountable owner | store=false does not replace ZDR/PSP approval |
OpenAI project allowlists take precedence over organization lists; the entries are not merged. They control requests sent to the OpenAI API, not user sign-in or an agent’s runtime network. Review who may call the model separately from where its tools may connect, so an inbound control does not conceal unrestricted tool egress.
OpenAI: API IP allowlist · ChatGPT Learn: Recommended configuration
{
"model": "gpt-6-cyber",
"access_programs": {
"cyber": "daybreak_red"
},
"safety_identifier": "demo-user-01",
"input": "For an owned offline example only, propose regression criteria for an owner-authorization fix. Do not call tools."
}4. CVP differs: granted workspaces, people and continuing compliance
CVP approval still requires provisioning appropriate to the surface. A Console organization assigns and activates a grant on a qualifying workspace. A paid subscription, a visible model name or organization membership does not establish that state. Before changing members, credentials or connected services, recheck the grant and Qualifications panel and record who approved the change.
Claude Help: Cyber Verification Program · Claude Help: Assign a program to Console workspaces
Individual applications currently support Defense on a paid plan only. The individual security requirements distinguish OAuth for interactive applications from WIF for programmatic workloads where offered. They specify phishing-resistant MFA and retirement of long-lived static credentials by 15 December, with storage, personal-use and at-least-every-7-day replacement conditions for interim keys. Existing customers may remain under current terms; do not automatically apply a new page to every historical grant.
Claude Help: CVP security requirements · Claude Help: Cyber Verification Program
For organizations, continuing controls also cover user attribution, revocation, offboarding and incident handling. Treating approval as a one-time milestone misses later configuration drift. WIF binds tokens to service accounts and workspaces and reduces long-lived key risk, but the upstream identity provider, workspace permissions and logging still need independent protection.
5. Data boundaries: what do ZDR, PSP and EFS actually govern?
ZDR does not mean no record exists anywhere. OpenAI PSP places encrypted safety records in customer-controlled storage for review in a protected automated runtime; the current guide describes a 30-day TTL. Confirm project status, permissions, lifecycle and operating responsibilities together. Adding a Boolean field in an application does not establish compliance.
| Scope | Current public boundary | Suggested operating practice |
|---|---|---|
| OpenAI ZDR+PSP | An applicable project connects and validates customer-controlled storage; automated safety review | Register storage and retention status with the task; minimize input |
| Individual CVP Defense | Traffic under the grant is retained and monitored; ZDR is unavailable | Submit only approved, redacted material |
| Claude Covered Models / EFS | Default retention differs from organization exceptions and phased EFS arrangements | Confirm the organization’s agreement and configuration without assuming general availability |
Claude Covered Models have a default minimum retention of 30 days, with investigation, legal and other exceptions. EFS arrangements and phased availability for eligible organizations are not a universal retention-off switch for CVP individuals. PSP and EFS both concern monitoring architecture, but their contracts, human-review arrangements and setup are not interchangeable. Ask security and compliance owners to confirm the actual project and model data path before choosing input detail.
Claude Help: Covered Models · Anthropic: Enterprise Frontier Safeguards · Claude Help: CVP security requirements
6. Give each task a versioned authorization scope manifest
The format below is our local record design, not an OpenAI or Anthropic API parameter. It is deliberately marked unapproved. Before actual execution, the asset owner must supply a valid approval reference, pinned commit and permitted files. While example values remain, the task stays in planning. A model cannot promote its own state to approved.
{
"schema": "defense-scope/v1",
"approval_ref": "DEMO-NOT-A-GRANT",
"state": "example_unapproved",
"repo": "toy-repo",
"commit": "DEMO_COMMIT",
"read_paths": [
"src/owner.py",
"tests/test_owner.py"
],
"network": "deny_all",
"tools": [
"read_file"
],
"output": "proposal_only",
"window_minutes": 30,
"human_gate": [
"writes",
"network",
"delivery"
]
}Separate readable files, permitted tools, output mode, time budget and human gates so a successor can see what a task did and did not do. Hash actual inputs and archive the manifest, tool configuration and outputs together. A demonstration commit name is not evidence of integrity.
This example permits only two synthetic files, denies all network access and produces proposals only. If further directories, connectors, dependency downloads or file writes become necessary, pause and explain the missing evidence to the owner before creating a new scope version. Instructions embedded in repository comments, websites or tool outputs cannot change authorization.
ChatGPT Learn: Recommended configuration · OpenAI: Guardrails and human review
7. An original delivery chain: an acceptable artifact at every step
This diagram describes our proposed transitions, not vendor response-state enums. Input versions, authorization, finding evidence and acceptance conditions must correspond before a candidate becomes “fixed in the synthetic fixture.” A real application still requires call-chain, deployment and maintainer review. Missing evidence may end a task as unresolved rather than forcing a success report.
- Register assets and approved scope; check the surface and data policy
- Pin synthetic inputs and versions; perform read-only analysis
- Record candidate evidence and unknown calling conditions
- Have a person review the minimal proposal; retain the original
- Compare before and after using identical regression criteria
- Hand over evidence, unresolved issues and review ownership
8. A synthetic finding: prove the observation before discussing risk
Consider an offline teaching function, can_read_invoice(actor, owner). Its rule permits access only when nonempty string identities exactly match. The original implementation returns only bool(actor). This supports a local observation that an owner comparison is missing; it does not establish a vulnerability in a real invoice endpoint. A caller might add checks, and this example has no route, database or deployment.
{
"schema": "finding/v1",
"id": "DEMO-F001",
"scope_ref": "DEMO-NOT-A-GRANT",
"file": "src/owner.py",
"symbol": "can_read_invoice",
"observed_code": "return bool(actor)",
"expected_rule": "nonempty_string_identity_equals_owner",
"status": "potential_finding",
"evidence_kind": "synthetic_source",
"unknowns": [
"call_site_checks",
"production_exposure"
],
"remediation": "propose_owner_comparison",
"acceptance_suite": "toy-owner-v1"
}Analyze only the supplied synthetic function and rule. Do not use tools or external systems.
Return: observed code, expected condition, direct evidence, missing call-chain information, minimal patch proposal and regression criteria.
Mark unsupported impact as unknown. Do not claim exploitation, a real-system fix or customer authorization.
If files outside scope are needed, list them and stop.Separate facts from inference in the finding record. File and symbol locate evidence; unknowns constrain the conclusion; the expected rule gives remediation a target. Do not copy a model’s severity rating directly into an external report. A maintainer must reconsider reachability, identity provenance and data sensitivity. Preserving disagreement is more useful than manufacturing a definitive score.
9. Turn a patch proposal into repeatable local acceptance
The proposed change compares the two identities after type and nonempty checks. The complete Python program below uses synthetic in-memory values only. It has no file reads, network or model SDK and can run in a separate local file. We ran only this offline fixture: 4 of 6 conditions failed before the change and 0 failed afterward. This is neither a model evaluation nor a real-system security conclusion.
import hashlib
import json
def before(actor, owner):
return bool(actor)
def after(actor, owner):
return (
type(actor) is str
and type(owner) is str
and bool(actor.strip())
and bool(owner.strip())
and actor == owner
)
cases = [
("demo-alice", "demo-alice", True),
("demo-alice", "demo-bob", False),
("", "demo-alice", False),
("demo-alice", "", False),
(1, 1, False),
(" ", " ", False),
]
expected = [case[2] for case in cases]
baseline = [before(a, b) for a, b, _ in cases]
patched = [after(a, b) for a, b, _ in cases]
assert baseline != expected
assert patched == expected
print(json.dumps({
"suite": "toy-owner-v1",
"cases": len(cases),
"before_failures": sum(a != b for a, b in zip(baseline, expected)),
"after_failures": sum(a != b for a, b in zip(patched, expected)),
"fixture_sha256": hashlib.sha256(
json.dumps(cases, separators=(",", ":")).encode()
).hexdigest(),
}, sort_keys=True))The program first asserts that the original violates the rule, then checks the revised implementation condition by condition. This prevents an empty test set from making both sides appear successful. Its digest pins the example inputs. A real project also needs source-file hashes, a patch diff, tool versions and original results; never attach this teaching fixture’s hash to different code.
Zero failures establish only these 6 conditions. Case handling, identity normalization, inherited permissions, service identities and the call chain are outside this example. A real maintainer must add applicable cases. Once a proposal enters a project, check legitimate behavior, add project tests and use the established deployment approval process rather than letting a model deploy it directly.
10. Maintainer handoff: explain results, boundaries and cost together
A usable handoff includes scope reference, input version, finding ID, original implementation, proposed change, retest procedure, machine results and unresolved questions. A “fixed” statement must identify whether it was confirmed in a local function, a complete application or a deployed version. These are different claims. Keep raw customer data and tokens out of the packet.
Using only the supplied scope-manifest, finding and offline results, draft a maintainer handoff.
Separate observed facts, inferences, unverified conditions, proposed changes, regression results, residual risks, owner and next step.
Attribute “6 conditions, 4 original failures, 0 revised failures” only to toy-owner-v1.
If asset approval, real source versions or deployment evidence are missing, mark the work incomplete. Never invent proof.Tie cost to the task as well. At the current GPT-6-Cyber standard short-context prices of $20 per million input tokens and $100 per million output tokens, an assumed 10,000 billed input and 2,000 output tokens gives 10000/1000000×20 + 2000/1000000×100 = $0.40. Two identical runs give $0.80. This is an unexecuted illustration excluding caching, tools, review, retries, storage and other charges; actual rules and measured usage determine the bill.
Track maintainer acceptance, unsupported findings, rework causes and the evidence needed for closure. We did not measure these metrics. They help compare a later model and workflow rather than promise that a higher access tier is cheaper. Establish repeatable low-risk delivery before expanding resources and call volume.
11. Continuing operations: distinguish failures from analytical conclusions
After the local fixture passes, a real workflow can still stop because of grant, data, network or review failures. Preserve the exact error category and redacted time. Do not interpret every 403 as failed identity verification, or bypass approval because a model answered. OpenAI API safeguards can still trigger cyber_policy; approved access does not remove those checks.
| Local state | Evidence to retain | Next action |
|---|---|---|
| access_unknown / blocked | Project/workspace, redacted error, actual model and control status | Ask an administrator to check approval; stop automatic retries that seek broader access |
| potential_finding | Source location, expected rule and unknown calling conditions | Ask the maintainer for more evidence |
| fixed_in_fixture | Pinned inputs, before/after results and applicability limits | Continue application review without claiming deployment |
| inconclusive / control_failed | Missing information or failed logging, permissions or review | Stop, preserve artifacts and repair the control |
Anthropic’s IP allowlist beta, announced on 5 November, adds controls for covered platform entry points. It does not control Managed Agents runtime networking or stand in for every third-party cloud policy. Review entry configuration separately from the execution environment. Role changes, token expiry, storage validation and model-mapping changes can all trigger review without waiting for an incident.
Claude Platform: Release notes · Claude Platform: IP allowlist scope
12. When can this expand into a team or customer service?
CVP currently accepts managed-service applications: the provider’s own staff operate the model, customers receive final deliverables, and each service needs separate approval. A product that customers can trigger or configure remains a product, for which only expressions of interest are currently accepted. Customer testing authorization does not replace vendor service approval. Daybreak’s public enterprise guidance likewise limits access to approved internal work; do not assume individual or internal access covers an external product.
Claude Help: Cyber Verification Program · ChatGPT Learn: Models and Trusted Access
Start an upgrade discussion with the actual gap. Is the current model unable to perform the analysis, or does the team lack maintainer review, data agreements or isolation? A more appropriate approved model may address the first problem; process and controls address the second. A seller page, subscription name or verification service cannot guarantee a particular grant, model, project or customer-delivery permission.
We reread official documentation and retained public snapshots, hashes and readable differences. We did not test GPT-6-Cyber, apply to CVP, access accounts or assess real systems. Individual eligibility, capacity, latency, private terms and final approval remain matters for the relevant official channel. Establish delivery habits on this network-free synthetic case before admitting approved real tasks into your change process.