Claude Writes 80% of the Code Now, Here's How Anthropic Keeps It Secure!

Hey everyone, it's Shiichan! Today I want to talk about how Anthropic secures its own development process. AI is writing more and more code these days, and I've always wondered how teams actually keep things secure when that happens.

Claude Blog claude.com

What was announced?

Anthropic published a post on the Claude Blog explaining how it protects its internal software development lifecycle (SDLC). It's written by Jason Clinton, Anthropic's Deputy CISO. Claude now authors about 80% of the code merged into Anthropic's codebase, and engineers ship 8x as much code per quarter on average compared to 2021-2025. Because development speed has grown so much, Anthropic had to rebuild its security approach from the ground up.

The story so far

Once developers adopted agentic coding tools, the CI review stage became a bottleneck. Pull requests were being produced faster than human reviewers alone could keep up with.

What changes

Anthropic built AI agents into every stage of development, Plan, Code, Test (CI), Deploy (CD), Monitor, and Governance, to raise both the quality and volume of reviews. The share of PRs receiving substantive review comments jumped from 16% to 54%. The key idea wasn't relying on one all-purpose agent, but running multiple specialized agents in parallel and having them validate each other's findings. Looking back at past incidents, Anthropic found that roughly one-third of them would have been caught by this current automated process. For any engineering org rolling out AI coding tools at scale, this is a practical playbook worth studying.

Dive Deep

The post breaks the development lifecycle into six stages:

• Plan: Security review is built into the design stage instead of being an afterthought
• Code: Security guidelines live in CLAUDE.md files, paired with a dedicated plugin that flags issues in real time as code is generated. Bugs that get found feed back into the guidelines, creating a closed loop that prevents the same mistake from happening again
• Test (CI): Multiple specialized review agents run in parallel and validate their own findings, which pushed the share of PRs with substantive review comments from 16% to 54%
• Deploy (CD): Continuous AI-powered DAST scans run in staging, targeting system-level vulnerabilities that span multiple services. Claude found and helped fix over 500 high-severity open-source vulnerabilities in February alone
• Monitor: Alert-triage agents review production logs and write postmortems, but they can't deploy fixes themselves. Deployment always requires human approval, keeping the blast radius contained
• Governance: Codebases are tiered by risk level, new AI reviewers start in shadow mode, automated approvals get sampled for review, and every agent action is logged to a SIEM for later auditing

The post also mentions Intercom as an outside example: after auto-approving 19% of PRs, they doubled their deployment frequency while cutting downtime by 35%. So this isn't just an Anthropic-specific shift, it looks like a broader industry trend.

Wrap-up

• Claude now writes about 80% of the code merged at Anthropic, and shipping volume is up 8x per quarter compared to 2021-2025
• The share of PRs with substantive review comments rose from 16% to 54%, and roughly a third of past incidents would be caught by the current automated process
• AI agents are embedded across Plan, Code, Test, Deploy, Monitor, and Governance, while critical actions like deployment stay gated behind human approval and SIEM logging
• The security team's job is shifting from chasing individual bugs to overseeing the loops and dashboards that catch them

If you're a security engineer or engineering manager rolling out AI coding tools at scale, this one's worth a close read!