Cortex brings source analysis, optional runtime validation, AI-assisted remediation, and revision-specific evidence into one governed repository workflow.
A source scan flags a potential access-control issue. A developer reviews a proposed fix, and the pull request passes its checks. Then another commit updates a dependency before release. Which security evidence still applies to the code the team is about to ship?
The scan, the proposed change, and the passing checks each answer part of that question. They can also describe different versions of the same software. It usually falls to the people responsible for shipping to work out which evidence belongs to the release candidate and where a fresh check is needed.
Cortex AI Repository Defense gives teams a way to answer that question. Start with a quick scan of an authorized repository, add runtime coverage where supported, and bring independent verification to eligible repairs. The assessment records what was examined, which evidence applies to the specific revision, and which decisions still need review. That gives developers, security reviewers, and release owners a common record from the first finding through the release decision.
What changes when the revision becomes the unit of assurance
AI Repo Defense associates a campaign with a specific repository revision. A campaign is the assessment record for that revision, including the scope, progress, findings, and supporting evidence. The branch name describes where the team is working. The revision identifies the source that was assessed. When a relevant input changes, the earlier report remains useful history and the new decision requires current evidence. When authorized work repeats against the same revision and equivalent coverage, compatible evidence can be reused and reported as such.
This matters beyond application source. Dependencies, build inputs, and infrastructure-as-code or deployment configuration can change the conditions under which a piece of code is exposed. Where the selected analysis supports those materials, they can inform the review. A repository assessment describes the material and behavior it actually covered, with deployment-specific questions left open when they require separate evidence.
| Capability | What it contributes | Question it helps answer |
|---|---|---|
| Static security scanning | Source-level observations about potential weaknesses | Where should a reviewer look in the current code? |
| AI-assisted code generation | A proposed edit and an explanation for review | What change might address the issue? |
| Runtime validation | Observations from an approved test environment | What behavior was exercised under the selected conditions? |
| Pull-request verification | Checks and review evidence for a proposed revision | What passed on the revision being considered for merge? |
| Cortex AI Repository Defense | A revision-bound record connecting coverage, findings, review, and eligible repair evidence | What do we know about this revision, what action is supported, and what remains open? |
Each capability retains its own job. A static scanner can surface an issue even when a project has no runnable service. A coding assistant can shorten the work of drafting a fix. Runtime validation adds observations about behavior. AI Repo Defense carries those results into a security decision with the context needed to interpret them.
Start with source. Add runtime evidence where it answers a question
The default is a quick static examination. It gives a maintainer a practical first pass and a useful result for repositories such as libraries, tooling, and configuration projects where running a service may add little. Teams can choose deeper source analysis or runtime coverage for a supported project when the decision warrants it. A repository that cannot run in a supported test environment still has a meaningful source assessment, with runtime coverage shown separately.
The initial qualification exercised four optional runtime coverage categories. Each addresses a different question about a supported test deployment.
| Runtime coverage | What a reviewer can learn |
|---|---|
| Scoped discovery | Which application surfaces were observable within the approved test scope? |
| Passive web analysis | What security-relevant characteristics were visible in the application’s responses? |
| Curated HTTP checks | Which selected security expectations were checked at the web interface, and what was observed? |
| API behavior coverage | Did the measured behavior of supported API operations match the selected expectations? |
Coverage and findings are two different measurements. Coverage records which requested categories produced qualifying evidence. Findings record the issues surfaced by the assessment. An assessment can complete with zero findings for the checks that ran. An unavailable runtime check is a coverage gap and remains visible. Even four of four requested categories describe the selected assessment scope, with behavior beyond that scope still unevaluated.
What makes evidence reusable
A repeated request can mean “show me what we already know about this commit” or “exercise a different environment or scope.” Reuse is appropriate when the authorized project, revision, scope, and relevant assessment conditions still match. For example, an existing source report may remain applicable when a team requests runtime coverage for the same revision. The newly requested scope still needs its own applicable runtime evidence.
The distinction matters in a review meeting. A result marked executed describes new work. A result marked reused brings forward applicable evidence from earlier work. Both can support a decision. The label lets a reviewer see where the result came from and decide whether the task calls for a new observation.
A finding is a security claim with a level of support
Severity is one input to triage. Reviewers also need to know whether a component is exposed in the relevant context, what conditions the finding depends on, and whether the observed behavior supports the concern. Several alerts can describe different manifestations of the same underlying issue. Treating each as an independent repair task can waste time and leave another affected location unresolved.
AI can assist with interpreting related signals, summarizing context for review, and drafting a bounded repair proposal where enabled. The strength of the conclusion follows the evidence. A static observation can warrant action on its own, particularly when the code clearly violates a policy. Runtime validation can support a claim, show that it does not hold under the assessed conditions, or leave it unresolved. Customer permissions and policy determine what may be accessed, run, changed, or shared.
| Finding, coverage, or repair status | What it communicates |
|---|---|
| Static observation | Source analysis surfaced a concern for review. Its runtime behavior may still be an open question. |
| Supported by validation | A bounded check supplied additional evidence for the claim under recorded conditions. |
| Disproved under the assessed conditions | Evidence contradicted the specific claim within the assessed scope. |
| Inconclusive | A check produced insufficient evidence to decide the claim. |
| Coverage unavailable | The selected check could not be completed or its evidence was unavailable. |
| Repair verified | The required checks support the repair for the specific revision reviewed. |
A published severity score is useful for comparing known issues. Prioritization also benefits from available evidence about exploitation, exposure, and impact in the affected software. A team may give an exposed weakness with known exploitation precedence over a higher-scoring issue in a less exposed component. For a finding with limited public information, the review can explain urgency using repository-specific observations and keep unknowns visible. When several findings share a root cause, reviewing them together also helps the team decide whether the proposed repair covers every affected location identified in scope.
A plausible patch and a verified repair answer different questions
A patch is a proposal about what should change. Verification is an assessment of what did change and how the affected software behaved afterward. A small diff can have a wide effect if multiple components depend on the behavior. A large diff can make it harder to tell whether the security issue was addressed or merely moved.
For eligible remediation workflows, AI Repo Defense can connect the original finding, affected revision, proposed repair, and independent validation. Where supported, reviewers can compare the relevant behavior before and after the change. They can then check whether the security expectation holds, whether required regression checks passed, and whether the evidence belongs to the pull-request revision being considered. Missing required evidence leaves the repair unverified. When the pull request changes, the new revision needs evidence for the decision being made.
Pull-request verification fits naturally into that process. It contributes evidence about the proposed change. Repository Defense provides the context for why that change matters to a finding and whether the applicable security checks support closing it. Proposal, verification, and authorization remain distinct decisions. An authorized reader can use Repository Defense for reporting, with repair and release actions governed by the controls that apply to the project.
Those distinctions become especially important across a portfolio. A developer needs enough context to work on a finding. A security lead needs to see which projects have current evidence and where coverage is incomplete. A release owner needs to know what passed for the candidate and which actions still require approval. A consistent assessment record lets those roles work from the same facts while keeping access to sensitive evidence and authority to act separate.
What this means for the release in our example
For the access-control finding, the security reviewer needs evidence that the expected restrictions hold and that permitted use still works. For a supported service, an approved test environment can supply that evidence. The developer has the finding and the proposed change in view, while the release owner can see which checks apply to the candidate that includes the dependency update.
An unavailable check leaves the finding and its coverage gap visible. A verified repair gives the team a record to review at closure, including the affected revision, the validation result, and any remaining limitations. Where revision monitoring is enabled, a later material change can prompt reassessment. This gives maintainers a way to revisit the decision as the software changes.
What we observed on one revision
In our engineering evaluation, an authorized test project produced two reports for the same revision. The first requested scoped discovery and recorded one category executed. A later repeat assessment requested all four runtime categories and carried forward eligible evidence for all four. The repeat report recorded 2 findings and 6 evidence artifacts.
Runtime coverage reported for the same source revision
Discovery Web HTTP API Covered
Scoped assessment Executed — — — 1 of 1
Repeat broader assessment Reused Reused Reused Reused 4 of 4
— = category outside the requested scope
Reused = applicable evidence carried forward from prior authorized workThe repeat assessment demonstrates that existing runtime evidence can carry forward into a new report and remain identifiable as reused. The findings and artifacts are totals for that report. A paired static-only run would be needed to measure additional findings attributable to runtime coverage. These results describe one project and do not establish detection accuracy, scan speed, or remediation success across repositories.
What we checked before qualifying the release
The figures below reflect the checks exercised and the acceptance scope for this release candidate. Future qualifications may cover a different set of checks, phases, and client surfaces.
| Observation | Recorded result | What the result establishes |
|---|---|---|
| Focused Repository Defense checks | 224 core checks and 35 runtime checks passed for this qualification | Both focused suites passed within the scope exercised for this release. |
| Release acceptance | 19 qualification phases and 5 client surfaces covered | Acceptance exercised the shared behavior across VS Code, the browser extension, mobile, CLI, and Cortex Discover. Available actions vary by client and role. |
| Live release qualification | Passed | The final gate accepted the live service readiness result, verified the authorized provider revision, and accepted fresh authenticated approval evidence. |
These checks qualify the release candidate for the capabilities covered by the gate. Repair verification and revision monitoring remain subject to the applicable project policy and capability availability. Customer outcomes call for different measurements, including time to a reviewable decision, repeat execution avoided, and the share of proposed repairs that pass independent validation. Comparing those outcomes across tools requires a defined baseline and comparable repositories, scopes, and revisions.
Repository defense and collective cyber defense
Software supply chains connect small maintainers, enterprises, service providers, and public institutions. A project may be developed by a small team and used by systems that team has never seen. When a maintainer can examine an authorized project, produce a repair with clear evidence, and communicate the affected versions responsibly, downstream users have a better basis for deciding how to respond.
Pervaziv AI joined the Open Secure AI Alliance, founded by NVIDIA and other industry participants and now hosted by the Linux Foundation. The alliance brings organizations together to develop open security tools, research, and shared defenses for AI systems. Repository Defense applies those principles to authorized defensive work with clear scope, visibility into what happened, and evidence that authorized people can review. Pervaziv AI’s Open Secure AI Alliance announcement · The Linux Foundation’s announcement
Pervaziv AI is also a signatory to OpenAI’s Call for Collective Action on Cyber Defense. That initiative calls for broadening access to AI-powered defense and improving coordination across organizations. We are applying that commitment by helping maintainers and engineering teams find weaknesses earlier in authorized codebases, validate repairs, and give the people responsible for releases a clear basis for action. Pervaziv AI’s collective cyber defense announcement · OpenAI’s open letter and signatories.
The two efforts address complementary responsibilities. One focuses on securing the AI and infrastructure used to do the work. The other focuses on getting useful defensive capabilities to more people. At the repository level, that means lowering the barrier to the first scan while keeping authority, evidence, and responsible disclosure in the hands of the people accountable for the software.
Start with a codebase you’re authorized to assess
Cortex has a freemium entry point. Use the Free Tier for no-cost file scanning or an eligible free trial to scan an authorized codebase, including an open-source project you maintain or are authorized to assess. Review the findings and coverage from a first source scan, then decide whether the project warrants deeper analysis. Teams can upgrade for larger repositories, broader coverage, or organization-level controls. Trial eligibility, repository limits, and advanced coverage depend on the plan. Cortex pricing and plans.
The first scan gives you a place to look. A revision-specific security record lets the next person see what was checked, what the evidence supports, and what still needs work. That record is what lets defensive work carry forward from the first finding to a release decision.


