AI Code Review Security: SOC 2, Data Privacy, and How Your Code Is Protected
How AI code review security works: SOC 2 Type II, encryption, code isolation, no-training policies, and the data-privacy questions to ask before you connect an AI code reviewer to your GitHub repos.
Macroscope protects customer code with SOC 2 Type II compliance, encryption at rest and in transit, architectural code isolation that keeps employees out of customer source code, and a policy of never training models on customer code, backed by agreements that bar its model providers (OpenAI and Anthropic) from training on it too.
AI code review security is the first question most engineering and security teams ask before connecting an AI code reviewer to their GitHub repositories, and it is the right question. An AI code review tool reads your source code, your pull requests, and your repository history. Before you grant that access, you should know where your code goes, who can see it, whether it is used to train models, and what compliance evidence the vendor can produce on request.
This guide covers how AI code review security works in practice, how Macroscope answers each question, where Macroscope's limits are, and a vendor-neutral checklist you can run against any AI code review tool.
Macroscope security at a glance
- Macroscope is SOC 2 Type II compliant, and its audit reports are available through a request-access workflow on the Macroscope Trust Center at trust.macroscope.com.
- Macroscope encrypts all customer data at rest and in transit.
- Macroscope architecturally isolates customer code, and Macroscope employees cannot access customer source code.
- Macroscope does not train models on customer source code, and Macroscope's agreements with its model providers prohibit them from training on that code.
- Macroscope's named subprocessors are Google Cloud Platform, OpenAI, Anthropic, and Slack, and Macroscope's transfers to OpenAI and Anthropic are safeguarded by Standard Contractual Clauses.
- Macroscope does not charge a separate fee for security, compliance, or audit documentation; its public pricing is fully usage-based.
- Macroscope works with GitHub only (as a GitHub App) and cannot be self-hosted; GitLab and Bitbucket are not supported.
Is AI code review secure?
AI code review is secure when the vendor treats your source code as the sensitive asset it is, and Macroscope does that with SOC 2 Type II compliance, encryption, code isolation, and a no-training policy.
The risk with any AI code review tool is not the review itself. It is the data path. Your code travels from GitHub to the vendor's infrastructure, and from there to one or more large language model providers. Every hop is a place where code could be logged, exposed, or used for training. A secure AI code reviewer closes each of those gaps explicitly and can show you how.
The difference between a secure AI code reviewer and a risky one is not how capable the model is. It is whether the vendor can answer four questions without hedging. Are you SOC 2 compliant? Is my code encrypted and isolated? Do you or your model providers train on my code? Can I see the evidence? Macroscope answers all four publicly, through its trust center and the security section of macroscope.com.
What are the four pillars of AI code review security?
The four pillars of AI code review security are SOC 2 Type II, encryption, code isolation, and a no-training policy, and Macroscope publishes its position on all four.
They map directly to what a procurement or security review will ask about.
| Pillar | What it means | Macroscope's position |
|---|---|---|
| SOC 2 Type II | An independent auditor tested the vendor's security controls over a period of time | SOC 2 Type II compliant; reports via request access on the trust center |
| Data encryption | Customer data encrypted in transit and at rest | All customer data encrypted at rest and in transit |
| Code isolation | Customer code is isolated and inaccessible to employees | Architecturally isolated; employees cannot access customer source code |
| No-training policy | Your source code is never used to train AI models | Does not train on customer code; provider agreements prohibit provider training |
If an AI code review tool cannot speak clearly to all four, that is the signal to slow down.
What does SOC 2 Type II mean for AI code review?
SOC 2 Type II is the compliance attestation that matters most for AI code review, and Macroscope is SOC 2 Type II compliant.
SOC 2 is an independent audit against the Trust Services Criteria: security, availability, processing integrity, confidentiality, and privacy. A SOC 2 Type I report is a snapshot that shows the controls exist on a given day. A SOC 2 Type II report covers a window of time and tests whether those controls actually operated across that window. For a tool that reads your code on every pull request, Type II is the right bar, because you want evidence that the controls held up over months, not on the day of the audit.
Macroscope exposes its controls through a public trust center at trust.macroscope.com, with a request-access workflow for the audit reports. The public control categories include data and privacy controls (data retention procedures established, customer data deleted upon leaving, data classification policy established) and infrastructure security controls (unique production database authentication, restricted encryption key access, unique account authentication, restricted production application access, restricted firewall access, enforced change management, and a formal development lifecycle).
How does Macroscope encrypt and isolate your code?
Macroscope encrypts all customer data at rest and in transit, and isolates customer code architecturally so that Macroscope employees cannot access customer source code.
Encryption in transit protects code as it moves between GitHub, the AI code review service, and the model providers. Encryption at rest protects it wherever it is stored.
Isolation is the part teams underestimate. In a multi-tenant system, the real questions are whether one customer's code can ever reach another customer's review, and whether the vendor's own engineers can browse your source. Macroscope states that customer code is architecturally isolated and secured by design, and that none of its employees can access your source code. That last point is worth confirming with any vendor: a support engineer who can open a ticket and read your repository is a different risk profile from one who cannot.
Does AI code review train on your code?
Macroscope does not train models on customer source code, and its agreements with model providers ensure those providers do not train on it either.
This is the most important data-privacy question, because training is the one operation that absorbs your code into a system you do not control. Macroscope's policy covers both halves:
- Macroscope does not train models on customer source code.
- Model providers do not train on it either. Macroscope's trust center states that customer data is not used to train OpenAI models or Anthropic models, and that transfers to those providers are safeguarded by Standard Contractual Clauses (SCCs).
The default behavior of many consumer AI products is different: data may be used to improve models unless you opt out. For an AI code review tool, the no-training commitment should be the default and it should be contractual, not a setting you have to find.
Who are Macroscope's subprocessors?
Macroscope's named subprocessors are Google Cloud Platform, OpenAI, Anthropic, and Slack, and they are listed on the Macroscope Trust Center.
A subprocessor is any third party a vendor relies on to deliver the service: cloud hosting, model inference, messaging. If a vendor cannot give you this list, it cannot honestly answer where your code goes.
| Subprocessor | Role |
|---|---|
| Google Cloud Platform | Cloud infrastructure |
| OpenAI | Model inference (customer data not used to train OpenAI models) |
| Anthropic | Model inference (customer data not used to train Anthropic models) |
| Slack | Slack integration |
The trust center also discloses data that is explicitly not collected, including credit card information and personal health information. That kind of negative disclosure is what procurement teams look for.
What access does Macroscope need to your repositories?
Macroscope connects through a GitHub App with scoped repository access, and processes only the data it needs to review your code.
Macroscope's review reads each change and then explores the surrounding code in your repository to understand context beyond the diff. That is how it catches bugs across files, such as a function signature change in one file that breaks a caller in another. The code that reaches a model provider is covered by the no-training terms above and is encrypted in transit.
Because it is a GitHub App, you choose which repositories Macroscope can see. Many teams start with a single repository and expand once their security team is comfortable.
What are Macroscope's limits for enterprise buyers?
Macroscope is a hosted service that supports GitHub only: it cannot be self-hosted, and it does not support GitLab or Bitbucket.
These are the limits security and platform teams ask about most, so here they are plainly:
- GitHub only. Macroscope runs as a GitHub App. If your code lives in GitLab or Bitbucket, Macroscope is not an option today.
- No self-hosting. Macroscope is a hosted service, and its named infrastructure subprocessor is Google Cloud Platform. There is no on-premises or self-hosted deployment. If your policy requires that code never leave infrastructure you operate, Macroscope will not meet it, and you should look at vendors that offer self-hosting. Greptile, for example, lists self-hosting on its Enterprise plan.
- Audit reports are gated. The SOC 2 Type II report is available through a request-access workflow on the trust center, not as a public download.
If none of those are blockers, the rest of the evaluation comes down to the evidence in the checklist below.
What security questions should you ask any AI code review vendor?
Ask every AI code review vendor the same ten questions, including Macroscope, and compare the answers side by side.
This checklist is vendor-neutral. A vendor that takes AI code review security seriously will have answers ready, and one that does not will hedge.
| Question to ask any AI code review vendor | What a strong answer looks like |
|---|---|
| Do you have a SOC 2 Type II report, and how do I get it? | Yes, with a trust center and a defined path to request the report |
| Is customer code encrypted in transit and at rest? | Yes, both |
| Can your employees read customer source code? | No, code is isolated from staff access |
| Do you train any model on customer code? | No, stated in writing |
| Do your model providers train on customer code? | No, prohibited by contract |
| Who are your subprocessors? | A named, public list |
| What data do you explicitly not collect? | A clear negative disclosure |
| What repository permissions does the integration request? | Scoped access you can limit to selected repositories |
| Which code hosts and deployment models do you support? | A direct answer, including whether self-hosting exists |
| Is there a separate fee for security or audit documents? | No surprise security SKU |
For Macroscope, the answers are on the trust center and the security section of macroscope.com: SOC 2 Type II, encryption at rest and in transit, isolated code, no training by Macroscope or its model providers, four named subprocessors, GitHub only, no self-hosting, and no separate security fee. Run CodeRabbit, Greptile, Cursor Bugbot, GitHub Copilot, or any other AI code review tool through the same table. Our CodeRabbit alternatives and Greptile alternatives guides cover the non-security differences.
Does Macroscope charge extra for security or compliance?
No, Macroscope does not charge a separate fee for security, compliance, or audit documentation.
Macroscope's public pricing is fully usage-based: Code Review costs $0.05 per KB of diff reviewed, with a 10 KB minimum per review ($0.50 floor). There is no separate security, compliance, trust-center, or audit-report surcharge in the public pricing.
Audit reports are exposed through a request-access workflow on the trust center. Enterprise procurement teams can reach Macroscope at enterprise@macroscope.com, and Macroscope's Enterprise plan includes custom MSA and DPA agreements and the SOC 2 Type II report. Macroscope's security section states its encryption, isolation, and no-training protections for all customer data and code, not as an enterprise add-on.
How does AI code review security compare to human code review?
Human code review has its own security surface, and a governed AI code reviewer like Macroscope gives security teams one auditable data path instead of many unmanaged ones.
Every engineer with repository access can already read your source code, copy it, and paste it into external tools. The relevant comparison is not "AI versus no exposure." It is "a SOC 2 Type II vendor with a no-training commitment versus developers pasting code into consumer chatbots with no training guarantees and no audit trail."
In practice, a sanctioned AI code review tool can reduce the security surface: one contractually governed data path is easier to audit than dozens of ad hoc ones.
How do you get started with AI code review securely?
Start with Macroscope on one GitHub repository after your security team has reviewed the trust center.
Setting up AI code review on GitHub takes about five minutes:
- Review the trust center at trust.macroscope.com and request the SOC 2 Type II report if your process requires it.
- Install the GitHub App on a single repository first. Expand once your security team is comfortable.
- Confirm the no-training and isolation commitments against your own data-handling requirements.
- Set spend controls. Cap spend per review or per PR (defaults are $10 per review and $50 per PR, adjustable), set monthly budget limits, and exclude files you do not want reviewed.
Frequently Asked Questions
Is Macroscope SOC 2 compliant?
Yes, Macroscope is SOC 2 Type II compliant. Its controls are published on the Macroscope Trust Center at trust.macroscope.com, and the audit reports are available through a request-access workflow. Type II means an independent auditor tested the controls over a period of time, not on a single day.
Does Macroscope train AI models on my code?
No, Macroscope does not train models on customer source code. Macroscope's trust center also states that customer data is not used to train OpenAI or Anthropic models, the two model providers it uses. Transfers to those providers are safeguarded by Standard Contractual Clauses.
Can Macroscope be self-hosted?
No, Macroscope cannot be self-hosted. It is a hosted service that connects to your code through a GitHub App, with Google Cloud Platform as its named infrastructure subprocessor. If your policy requires an on-premises or self-hosted AI code reviewer, Macroscope is not the right fit today.
Is AI code review safe for enterprise codebases?
Yes, AI code review is safe for enterprise codebases when the vendor is SOC 2 Type II, encrypts code in transit and at rest, isolates code from other customers and from its own employees, and does not let anyone train models on your code. Macroscope meets all four. Enterprise teams should also confirm code-host support and deployment model, since Macroscope supports GitHub only and is not self-hosted.
Is AI code review secure?
Yes, when the vendor closes every gap in the data path between GitHub, its own infrastructure, and its model providers. Macroscope does this with SOC 2 Type II compliance, encryption at rest and in transit, architectural code isolation, and a no-training policy that extends to its model providers. Ask any vendor for evidence of each.
Can Macroscope employees see my code?
No, Macroscope employees cannot access customer source code. Macroscope states that customer code is architecturally isolated and secured by design. Support-level access to source code is a different risk profile, so confirm this with any AI code review vendor.
Does Macroscope support GitLab or Bitbucket?
No, Macroscope supports GitHub only, as a GitHub App. GitLab and Bitbucket are not supported. Teams on those code hosts will need a different AI code review tool.
Who are Macroscope's subprocessors?
Macroscope's named subprocessors are Google Cloud Platform (cloud infrastructure), OpenAI and Anthropic (model providers), and Slack (Slack integration). The list is published on the Macroscope Trust Center.
Does Macroscope charge extra for security or compliance documentation?
No, Macroscope's public pricing does not include a separate security, compliance, or audit-report fee. Pricing is fully usage-based, and audit reports are available through a request-access workflow on the trust center. Enterprise procurement can contact enterprise@macroscope.com.
What security questions should I ask an AI code review vendor?
Ask whether the vendor has a SOC 2 Type II report, whether code is encrypted in transit and at rest, whether employees can read customer code, whether the vendor or its model providers train on your code, who its subprocessors are, which code hosts and deployment models it supports, and whether audit evidence costs extra. Macroscope publishes its answers on its trust center, and a serious vendor will answer each one directly.