50+ Features Anthropic Shipped in 52 Days — What Security and Compliance Practitioners Should Know

If you opened Claude six months ago and have not been paying close attention since, the product you remember is not the product that exists today. Anthropic has shipped more features in the first quarter of 2026 than most enterprise software vendors ship in two years. Two new flagship models. A million-token context window. Computer Use. Persistent memory. Skills. Channels. Auto Mode. A Code Review system. Multi-agent orchestration. Cron-like scheduled jobs. Excel and PowerPoint integration. A 33-page spec turning Claude into something closer to an operating system than a chatbot.

Most coverage of these releases has been written for software developers. That makes sense — Claude Code is where many of the most interesting features land first. But security practitioners, compliance professionals, and IT leaders who are trying to figure out how to use AI for technical work have largely been left out of the conversation. The feature list is moving too fast to track unless it is your job to track it.

This is that list. Specifically, the features that matter if you are doing security assessments, compliance reviews, framework analysis, or any of the technical-but-not-quite-engineering work that fills the day for vCISOs and security leaders. I have tried to call out which features are useful for what, and I have flagged the ones that change how the work gets done — not just the ones that are new.

I will also cover what was in the leaked Claude Code source code at the end. That part is the most interesting.

The Models You Should Know About

Two model releases this year reset the baseline for what Claude can do.

Claude Opus 4.6 (released February 5) is the flagship. Opus 4.6 is the model you want for hard reasoning work — long-form policy analysis, complex framework comparisons, anything where you need the model to hold a lot in its head at once. It scores 90.2% on BigLaw Bench (legal reasoning) and leads the industry on financial due diligence tasks. For security work, the practical impact is that Opus 4.6 can hold an entire CIS benchmark, a NIST framework, and a client’s environment configuration in context simultaneously without losing the thread.

Claude Sonnet 4.6 is the workhorse. It is faster and cheaper than Opus, with most of the capability for routine work. If you are running automated assessments, generating draft reports, or building skills that need to run reliably at scale, Sonnet 4.6 is usually the right call.

Both models now support a 1 million token context window in beta. That is roughly 750,000 words. For comparison, the entire CIS M365 Foundations Benchmark is well under 200,000 tokens. You can fit the full benchmark, the client’s tenant configuration export, and three months of audit logs into a single conversation and still have room for analysis. This was not possible six months ago and it changes what is feasible to do in a single pass.

Productivity Features That Matter for Technical Work

These are the consumer-facing features that ended up being more useful than expected for compliance and security work.

Persistent memory rolled out to all users (including the free tier) in early March. Claude now remembers your name, communication style, writing preferences, and ongoing project context across separate conversations. For consultants juggling multiple clients, this means Claude can keep track of which framework you typically work in, which audience you write for, and what your last few engagements looked like — without you re-explaining at the start of every session.

Inline charts and diagrams. Claude can now generate interactive charts, tables, and visualizations directly inside its responses. For compliance reporting, this means you can ask for “a control coverage heatmap by section” and get a working visualization, not a description of one.

Excel and PowerPoint integration. The Claude for Excel and Claude for PowerPoint add-ins now share full conversation context. You can analyze a control inventory in Excel, ask Claude to draft executive summary slides in PowerPoint based on the same conversation, and the slides will reference the actual data in the spreadsheet. This is the workflow that turns “I have findings” into “I have a board deck” without manual translation.

Projects workspace with persistent context. The new Projects feature gives Claude persistent memory and style controls scoped to a specific project. For consultants working multiple engagements, this is the difference between starting every session with context-setting and just resuming work.

Custom instructions and skills. Anthropic has been quietly building out a skill system that lets you package context, tools, and workflows so Claude triggers them automatically. More on this below.

Claude Code: The Feature That Changes the Most

Claude Code is Anthropic’s terminal-based agentic environment. For security practitioners, this is where the most consequential features are landing.

Computer Use. Claude can now operate your computer directly — open files, run dev tools, click, type, and navigate to complete tasks. For security work, this means Claude can drive an admin portal to verify settings, navigate documentation sites, and interact with tools that do not have APIs. Everything still requires permission, but the friction of “I need to manually do this part” is mostly gone.

Auto Mode. Auto Mode is the middle ground between approving every Claude action and giving it unrestricted access. Claude executes safe actions automatically while routing risky ones through approval. It uses two safety checks — an input-layer prompt-injection probe and an output-layer transcript classifier — to decide what is safe. Auto Mode currently works with Sonnet 4.6 and Opus 4.6 and is rolling out to Enterprise and API users. For long-running assessment workflows, this cuts the number of manual approvals from hundreds to dozens.

Code Review. Claude Code now has a built-in Code Review system where a team of agents runs deep review on every pull request. For security teams, this is the closest thing yet to “automated security code review that actually works.” The agents check for common vulnerability patterns, missing test coverage, and architectural issues — and surface findings the way a senior reviewer would.

Multi-Agent Orchestration (“Claude Cowork”). Multiple Claude Code sessions can now share context and coordinate tasks. Instead of one Claude doing everything sequentially, you can run a team of Claudes — one researching, one writing, one reviewing — that talk to each other. For complex assessments that would otherwise need to be split across multiple manual workflows, this is the closest thing yet to having a small team of analysts working in parallel.

Scheduled Jobs (/loop). The /loop feature turns Claude Code into a background worker with cron-like scheduling. You can schedule Claude to run a task on an interval — daily compliance checks, weekly threat intelligence summaries, hourly drift monitoring. This is not just automation; it is a persistent agent that runs without you having to remember.

MCP (Model Context Protocol) Skills. MCP is the standard Anthropic created for Claude to talk to external systems. MCP connectors now work in single-turn print mode, not just multi-turn sessions, and the result size limit was raised to 500,000 characters per call. In practice, this means Claude can call out to Slack, Gmail, GitHub, Linear, your CIS benchmark database, your evidence management system — and pull large amounts of data into a single workflow.

File Reference Efficiency. When you reference files with @, Claude now processes them more efficiently. Raw string content is no longer JSON-escaped, which means less token overhead and the ability to reference more files without burning through context. For working with large compliance documents, this is a meaningful upgrade.

The CLAUDE.md Pattern

This is not technically a feature — it is a convention Anthropic documented and the community discovered. You can drop a file called CLAUDE.md into any project directory and Claude Code will read it automatically as context. Inside, you put project-specific instructions: how to run tests, what conventions to follow, what context Claude needs for this codebase.

The reason this is showing up everywhere is that it fixes the biggest weakness of AI coding assistants: they do not know your project. CLAUDE.md teaches Claude what your project is, every time you start a session, without having to re-explain. For security teams, the same pattern works for compliance work — you can put your assessment methodology, framework references, and reporting standards into a CLAUDE.md and Claude will follow them.

It is the highest-leverage 30 minutes you can spend if you are using Claude Code for any recurring work.

What Was in the Leaked Source Code

On March 31, Anthropic accidentally leaked the TypeScript source for Claude Code. Bloomberg, WSJ, and TechCrunch all covered it. The leak revealed not just current features but a roadmap of what is coming.

The unreleased features in the leaked source include:

• Capybara — an unreleased model
• Opus 4.7 and Sonnet 4.8 — next-generation models
• Proactive Mode — Claude initiates work without being prompted
• Coordinator Mode — multi-Claude orchestration as a first-class capability
• Adversarial Verification — Claude reviews its own work for flaws before delivering it
• Token Budgets — explicit cost controls per task
• Reversible Context Collapse — compressing context without losing the ability to recover details
• Persistent Job Templates — saved workflows you can re-run
• Bring Your Own Compute — running Claude on your own infrastructure
• 24/7 autonomous background agents — Claude that just runs
• Cron-based scheduling as a built-in primitive
• 85 commands in the source, including /init, /plan, and many that have not been documented

The pattern is clear: Claude Code is moving from “tool you prompt” to “agent that works.” Proactive mode means Claude starts tasks on its own. Background agents mean Claude continues working when you are not at your desk. Coordinator mode means multiple Claudes working in parallel without you orchestrating them manually.

For security and compliance practitioners, this is the same trajectory we have been watching with vibe coding more broadly: AI moving from assistant to autonomous collaborator. The governance question that applies to vibe coding applies just as strongly here. If Claude is going to run jobs autonomously, in the background, on your behalf — what controls govern what it can touch, what decisions it can make on its own, and how you audit what it did?

What This Means for How You Work

The pace is the story. Anthropic shipped 50+ meaningful features in 52 days. By the time you read this, more have shipped. Building expertise on a specific feature is less valuable than building the habit of checking what is new every two weeks.

A few practical takeaways:

Use Opus 4.6 for the hard stuff. Long-form analysis, complex framework comparisons, anything where you need the model to hold a lot in its head. For routine work, Sonnet 4.6 is faster and cheaper.

Set up CLAUDE.md for any recurring work. Thirty minutes of setup compounds across every future session. If you use Claude for assessments, compliance reviews, or report generation more than once a month, this is non-negotiable.

Try Auto Mode for long-running tasks. If you are running multi-step workflows and approving every action, Auto Mode will save you most of that time. The two-layer safety checks mean you are not giving up control entirely — you are filtering out the routine approvals so you can focus on the decisions that actually matter.

Watch the leaked roadmap. Proactive mode, background agents, and coordinator mode are coming. If you build workflows now, design them so that when these features ship, you can plug them in without rebuilding from scratch.

Govern the autonomy. Every new capability that lets Claude do more on its own is also a new question about what controls apply. Define what data Claude can access, what actions it can take without approval, what gets logged, and what triggers a human review. This is not optional anymore — it is the foundation for using these tools responsibly at scale.

---

Genesis builds practical workflows for security and compliance teams using Claude Code, custom skills, and the agentic-os architecture. We are publishing how we do it as part of the Claude Training Log series — behind-the-scenes posts on the actual workflows, configurations, and patterns that produce reliable results.

Contact us if you want to talk about how AI fits into your security or compliance practice.