We’ve shipped a security-guidance plugin for Claude Code that helps identify and fix vulnerabilities as you’re writing code. Available for all Claude Code users. Install from the plugin marketplace (/plugins).

It runs via hooks and reviews code on three levels: - On file edits: looks for risky patterns like commonly misused dangerous libraries - After model turns: reviews the full diff for harder-to-spot issues - On commits: reads surrounding code to validate vulnerabilities

We've been using the plugin extensively at Anthropic. Across our internal rollout and benchmarks, we’ve seen a 30-40% decrease in security-related comments on PRs opened using the plugin. The plugin serves as a lightweight first pass, catching issues before a full code review

You can add org-specific rules in a claude-security-guidance.md file. Drop it in your repo or distribute via MDM. The plugin enforces your policies alongside the built-in checks. Docs:

Tried the new security-guidance plugin. Asked it to add eval() and a dangerous regex. It immediately blocked both with proper OWASP explanations.

software development slowly becoming two ais arguing in the sidebar about whose bug it is

Reset plz im dying :x

Claude now in cyber security

security is finally solved just like software development

This is a good first pass for catching risky code Claude writes. But it is not supply chain prevention. If an agent downloads a bad package, I want that blocked before install, not politely reviewed after the diff. Useful plugin but Not a package firewall.

Hey, please address this issue

Catching vulns while you're writing the code, not after you've already shipped it to prod is really a good move but the scary part is how many teams skip the after part entirely and their live site just sits there exposed..

I do not think people realize how valuable this is

The strongest security feature may end up being feedback at the moment of creation, not after deployment

個人開発勢からすると難しいニュースなので調べてまとめてみました！

I think this should be generalized to a new type of subagent: background reviewers. Run many of them, each scoped to different small things.

the entire agent stack now treats frontier models like nom packages - bump the version, ship the PR, done

Why does it write vulnerable code in the first place? Can't good practices be forced through a harness?

Multi-modal superpower 🦸🏻‍♀️

Is this x10 speed? Dude Claude code has been too slow this days 🥲🥲🥲

Catching a JSON Parser-Disagreement Field Smuggling vulnerability in the background while you're still writing the code. That's not a plugin, that's a security engineer running in parallel.

Nice, this is exactly what devs need. Catching bugs while you're still typing saves so much headache later. Good move.

You should either identify and fix the Max 20x usage-limit leak. Going from 50% to 88% in one task is impressive. Almost as impressive as announcing “more capacity” while the product behaves like someone drilled a hole in the quota bucket. At some point, this stops looking like confusing messaging and starts looking dangerously close to consumer deception. 👇🏾 Open Letter to Anthropic May 28, 2026 Anthropic Leadership & Claude Team, I am a Max-tier customer. One month ago, I could comfortably run three to four heavy, parallel coding sessions without ever touching the five-hour wall. Today, two sessions are enough to hit it reliably, and for the first time, I am also hitting the weekly cap. On May 6, you announced that five-hour limits would be doubled, peak throttling removed, and that fresh compute would finally give serious users breathing room. What actually arrived feels more like a quiet, unannounced halving for anyone using Claude Code the way it is marketed: long-context agentic work, parallel exploration, and real refactors. The meter does not show anything close to doubled headroom. Multiple users are reporting the same display bugs, where the system insists the limit is reached while the dashboard still shows quota left. My own logs match the pattern. You sold doubled capacity. Paying customers are experiencing the opposite. That is not a minor execution hiccup. That is a public promise followed by measurable regression. At this point, it starts looking less like ambitious roadmap communication and more like creative accounting with customer attention spans. The market has already started voting. Significant portions of serious development work are moving to Cursor, Aider, and back to OpenAI environments. Not because the models there are dramatically better, but because the usage experience is at least predictable and does not require constant detective work to figure out whether the announced improvements actually exist in production. Claude remains one of the strongest reasoning models available. That achievement is real. It is now being steadily undermined by a limits regime that punishes exactly the heavy, high-value usage you claim to support, combined with announcements that do not survive contact with actual usage data. I am publishing this because private channels have not moved the needle, and because the gap between what was promised on May 6 and what is observable today is now large enough to be embarrassing. Either publish the actual before-and-after usage data for Max-tier agentic workloads, or stop announcing capacity increases that do not reach the people paying the highest prices. One of those two options would be the adult version of this situation. The current version is starting to look like a very expensive magic trick performed on the customers funding the show. Samir Max customer who used to be able to work without playing usage detective

Mythos 0.1

This right? 👀

Shouldn't this be "native" to Claude and not a plugin?

This looks interesting, thank you!

A security plugin that lives inside coding agent not a scanner you run after. Curious how false positive-heavy it is in practice though.

Finally, security that meets you while you're coding instead of yelling at you in PRs two weeks later. The 30-40% drop in security comments at Anthropic is the kind of dogfooding stat that actually moves the needle

Security feedback inside the coding loop is the real win.

Great plugin, it's much better that the security checks happen while code is being written and not after

The org-specific rules file is the interesting part imo Catching eval() is table stakes. The real win is when repeated PR comments stop living in reviewer memory and get promoted into repo policy before the next diff Otherwise the plugin helps once, but the org doesn't actually get safer

the best security check is one you never have to think about hooks make that possible. installing this now

Catching vulnerabilities during development saves significant rework. Is the goal to fully automate fixes, or to guide the developer to understand and implement the secure patterns themselves?

security plugin inside the coding agent is such an obvious unlock. the future is just agents fixing the mess they were about to create 😭

Pairing secure code with residential routing is exactly how to keep AI agents unblocked. 🛰️🫡