Before going deeper, here is the overall picture in plain words.

Claude Code is Anthropic’s tool for working on software with an AI agent that can read a codebase, run commands, and make changes. Claude Science is a separate workbench for researchers, built to bring together databases, analysis tools, and computing resources. Claude Security is a product for security teams that scans a codebase for vulnerabilities and proposes fixes. It began life as a feature called Claude Code Security and was later renamed and expanded.

Because the names are similar, people sometimes assume that Science and Security are modes inside Claude Code. That is not quite right. Claude Science is its own application. Claude Security is its own product with its own entry point, and it connects to Claude Code in specific ways, such as opening a finding in a Claude Code session to work on the fix. Details change quickly while these products are in beta, so this guide describes the ideas and points you to Anthropic’s current pages for specifics.

Understanding Claude Code first makes the other two easier to place.

Claude Code is an AI coding agent. Instead of only answering questions about code, it can look through a project’s files, run commands, write and edit code, and carry out multi-step tasks. Developers use it in a terminal, in a web experience, and in other places, and it asks for permission before taking certain actions. It is used for building features, fixing bugs, explaining unfamiliar code, and writing tests.

What matters for this guide is that Claude Code is the general-purpose engine for working on code, and that it is a place where fixes get made. Claude Security’s findings can be taken into Claude Code to develop a patch. Claude Science’s role is different, and it has its own home, which later chapters describe.

It aims to put a scientist’s scattered tools into one place.

Anthropic introduced Claude Science in mid 2026 as an AI workbench for scientists, in beta for paid plans. Scientists often move between many databases, analysis environments, and terminals, each with its own format and quirks. Claude Science is designed to bring these into one environment, where Claude can help search literature, query databases, write and run analysis code, manage computing jobs, and produce figures and manuscripts that can be refined.

Anthropic describes it as an application for scientists that can run on a researcher’s own machines, such as a laptop or Linux system, on a remote machine, or through a lab’s high-performance computing setup. Anthropic has also said that it is not a new model and not a more capable model for biology. It uses Claude’s existing models and adds a workflow around them.

The pattern is plan, review, run, and check.

Anthropic’s descriptions suggest a human-friendly rhythm. The researcher describes a goal. Claude proposes a plan. The researcher reviews that plan. Claude then carries it out, writing and running code, and, for heavy jobs, preparing the work to be submitted to the lab’s own compute resources, such as a cluster, or to an on-demand cloud compute service. Results come back as artifacts such as figures, tables, and draft text, together with the details of how they were made.

flowchart TD
    A["Researcher states a goal"] --> B["Claude drafts a plan"]
    B --> C["Researcher reviews
and adjusts the plan"]
    C --> D["Claude runs analyses
and prepares compute jobs"]
    D --> E["Results appear as figures,
tables, and drafts"]
    E --> F["A reviewer agent
and the researcher check them"]
    F --> G["Refine until ready,
with code and methods attached"]
        

A result you cannot reproduce is hard to trust.

A central idea in Claude Science is that results should be reproducible. Anthropic describes figures and manuscripts that arrive with the code that created them, the computing environment, a plain-language explanation of the method, and the conversation that led to them. This makes it easier for other researchers, or for the same researcher months later, to see exactly how a result was produced, and to repeat or challenge it.

This matters because AI-assisted analysis can look polished while hiding choices that affect the outcome. By keeping the work visible, a workbench invites checking instead of blind trust.

Research data is often large, sensitive, or both.

Anthropic states that Claude Science runs on a lab’s own infrastructure, so large or sensitive datasets do not have to leave the systems they are already on, and only the context needed for each step of an analysis is sent to Claude. This design is meant to reduce the need to move whole datasets around. It does not remove the need for judgment. The context sent for each step may still include content from your data, so you should know what is being shared.

Research groups should also follow their institution’s rules, ethics approvals, and data agreements. Data about people, patients, or protected subjects can have strict requirements, and some programs of research may not allow certain data to be used with external services at all. When unsure, ask your institution’s data or compliance contact before using any tool.

AI help in research needs more checking, not less.

Anthropic describes a reviewer agent within Claude Science that checks the work, together with human checks. A second AI reviewer can catch some mistakes, such as inconsistent steps or unsupported claims. It cannot replace a scientist’s judgment about whether a result makes sense, whether the method suits the question, or whether the data support the conclusion.

Good practice is to treat Claude’s analysis as a draft by a capable but fallible assistant. Read the code, check the assumptions, test results against known cases, confirm citations by opening the sources, and ask whether the conclusion is as strong as the evidence allows.

It suits researchers who work with data and code, and it has limits.

Claude Science is aimed at working scientists, especially in data-heavy areas such as genomics, proteomics, and chemistry, where analysis involves many databases and tools. It can help people who write analysis code as a routine part of their research, and it can save time on repetitive set-up. It is less relevant for someone who only needs a quick answer to a general question, for whom ordinary Claude chat or the Research feature may be enough.

It helps security teams find weaknesses in their own code and fix them.

A vulnerability is a weakness in software that an attacker could use to cause harm, such as stealing data or taking control of a system. Security teams face a constant challenge: there are many possible weaknesses and too few people to find and fix them all. Claude Security is Anthropic’s product for scanning a codebase for vulnerabilities, checking the findings, and suggesting patches for the team to review and approve.

Anthropic first introduced it as Claude Code Security in early 2026 as a limited preview, then launched it as Claude Security in public beta for Claude Enterprise customers. It is described as reasoning about code more like a human security researcher, tracing how data flows between files and components, instead of only matching known patterns. Availability and plan details have been changing, so check Anthropic’s current pages.

The flow is scan, check, rate, and propose a fix.

According to Anthropic’s descriptions, you connect a code repository and Claude scans it, reading the source, understanding how parts interact, and tracing data through the system. Claude then challenges its own results before showing them, to reduce false alarms. Each validated finding comes with a confidence rating, a severity, an explanation, and a suggested patch. A person reviews each finding and decides what to approve.

flowchart TD
    A["Connect a repository"] --> B["Claude scans and reasons
about the code"]
    B --> C["Claude challenges its own findings"]
    C --> D["Validated findings with
confidence and severity"]
    D --> E["Suggested patch for each finding"]
    E --> F{"Security team reviews"}
    F -- "Approve" --> G["Fix is developed and shipped"]
    F -- "Reject or change" --> H["Adjust the patch or dismiss"]
        

A suggested fix is a proposal, and the human review is the safeguard.

Anthropic emphasizes that nothing is applied without human approval. Findings and patches appear for the team to review. A good reviewer treats each finding with healthy skepticism. They check that the vulnerability is real in the context of how the system is actually deployed, that the severity makes sense, and that the proposed patch fixes the problem without breaking other behavior.

Claude Code is where many fixes get developed.

Anthropic describes a smooth path from finding to fixing. Because Claude Security is built on Claude Code, teams can review findings and iterate on fixes within tools they already use. A finding can be opened in a Claude Code session to work through the patch with context. Anthropic also describes a Claude Security plugin for Claude Code in beta, which lets a team scan, validate, and patch directly inside Claude Code while keeping code within their environment.

This means the connection between the two is practical. Claude Security is the place that surfaces and tracks the issues, and Claude Code is a place where an engineer, with Claude’s help, writes the repair. The usual safety habits for Claude Code still apply: review diffs, run tests, and rely on version control.

Security is not a one-time event, so the product has grown to support ongoing work.

Anthropic has described updates that add ongoing code scanning and automatic fix tracking, so teams can monitor vulnerabilities as software changes. It also describes project-level context, such as a knowledge base where teams document how the code is deployed and which security policies apply, a threat model generated from the code that teams can review and edit, and an API that connects Claude Security to existing workflows to manage projects, view findings, and resolve or reopen issues. These features can change in beta, so read the current notes.

The same skill that finds weaknesses could be misused, so rules matter.

Anthropic has stated that the capabilities that help defenders find and fix vulnerabilities could also help attackers, which is why it frames Claude Security as a defensive tool with human approval at the center. Use it only on code and systems you own or are explicitly authorized to test. Do not scan or probe other people’s systems without permission. Keep findings confidential until they are fixed, and follow your organization’s process for handling and disclosing vulnerabilities.

Also be thoughtful about the access you give. A scanner reads your source code, so connect only the repositories that need scanning, and follow your organization’s policy on who may see findings, since vulnerability details are sensitive.

A single table makes the differences easy to see.

Start from the job, not the name.

flowchart TD
    A["What is your job?"] --> B{"Building or changing software?"}
    B -- "Yes" --> C["Claude Code"]
    B -- "No" --> D{"Analyzing scientific data
with code and databases?"}
    D -- "Yes" --> E["Claude Science"]
    D -- "No" --> F{"Finding weaknesses in
your own code?"}
    F -- "Yes" --> G["Claude Security,
with Claude Code for the fix"]
    F -- "No" --> H["Everyday Claude chat
or Research"]
        

Both specialist products are new, so availability moves quickly.

At the time of writing, Anthropic describes Claude Science as in beta for paid plans, with discounted programs for some research labs, and describes Claude Security as in public beta for Claude Enterprise customers, with wider availability said to be planned. Platforms supported, entry points, and features such as the Claude Code plugin have also been changing. These facts may already be out of date when you read this.

The sensible approach is to check the official product pages and help center, ask your organization’s administrator what is enabled for you, and avoid planning a project around a feature that is still in beta without a fallback.

Familiar situations show how the tools fit together.

Specialist tools bring focus, and they ask for care.

These habits cause most problems with specialist tools.

Follow these steps for either specialist tool.

A few popular beliefs about these tools are only partly true.

Quick answers to the questions beginners ask most.

Not exactly. Claude Science is a separate workbench application. Claude Security is a separate product that connects to Claude Code, for example by opening findings there and through a plugin for Claude Code.

It is Anthropic’s AI workbench for scientists, which brings databases, analysis tools, and computing into one environment and produces reproducible figures and manuscripts.

No. Anthropic says it is not a new or more capable model for biology. It uses existing Claude models inside a workflow built for research.

Anthropic says the app runs on your own infrastructure and only the context needed for each step is sent to Claude. Still, check what is sent and follow your institution’s rules.

It is a product for security teams that scans a codebase for vulnerabilities, validates findings, rates them, and suggests patches that people review and approve. It was previously called Claude Code Security.

Anthropic says every patch requires human review and approval. Treat suggestions as proposals, and test them before shipping.

Findings can be opened in a Claude Code session to develop the fix, and a Claude Security plugin for Claude Code is described as available in beta.

Availability differs and is changing. At the time of writing, Science is described as in beta for paid plans and Security as in public beta for Enterprise. Check the official pages.

No. Use it only on code and systems you own or are explicitly authorized to assess.

Review plans and findings, read the code, test with known cases or your test suite, verify citations, and get an independent human check for important results.

On Anthropic’s official product pages and help center, since these products are new and change often.

Claude Code, Claude Science, and Claude Security are related tools with different jobs. Claude Code is the agent for building and changing software. Claude Science is a separate workbench that helps researchers bring data, tools, and computing together and produce reproducible results, running on the lab’s own infrastructure. Claude Security is a defensive product that scans code for vulnerabilities, validates its findings, and proposes patches for people to approve, with a close connection to Claude Code for developing the fixes. In every case, human review matters most: check plans and findings, read the code, test results, protect sensitive data, use security tools only where you are authorized, and recheck Anthropic’s pages because these beta products keep changing.