Claude Code安全测试：用Opus 5.5自动化攻击演练

Claude Code tip: before you ship a vibe-coded site, have Opus 5.5 𝗽𝗹𝗮𝘆 𝘁𝗵𝗲 𝗯𝗮𝗱 𝗴𝘂𝘆. Send out 3 subagents with three test accounts (A, B and an admin) to actually run 𝘁𝗵𝗲𝘀𝗲 𝟭𝟰 𝗮𝘁𝘁𝗮𝗰𝗸𝘀, collect proof, then fix everything in one pass:

Impersonate other users → Without logging in, take the public key from your frontend and query every table in the database directly. See what comes back → Logged in as A, swap the id in a request for B's and try to read, edit and delete B's data → When A edits their profile, add role: admin to the request, then try editing the role in their own account data directly → Use a password reset link once, then click it again

Get around limits in the UI → For disabled buttons and hidden options, skip the page and send the request directly → Once the free uses run out, check whether calling the API directly still works → With only a little credit left, fire 20 requests at once and see whether more succeed than the credit allows → Fill forms with very long text, special characters, negative numbers and empty values → Change the post-login redirect (?next=) to another site, and also try values starting with // and @

Come in from outside → Put <img src=x onerror=alert(1)> in a username, a comment or Markdown, then open the page in a browser and see if an alert pops up → If the app fetches URLs (importing links, building previews), give it http://localhost and http://169.254.169.254 and check the server logs for an actual outgoing request. Skip this if there's no such feature → For pages and API responses A has seen, request the same address from another browser as B or logged out, and see whether A's content shows up → Open your hosting platform's preview links without logging in to the platform → Hide an instruction in an uploaded document or web page that tells your AI feature to read another user's data or call a tool it shouldn't, and see if it does

Collect proof → Every item has to actually run. If it was only read in the code and never run, it counts as "not proven" → After a fix, temporarily turn the protection off in local code and run it again. The attack must succeed. Turn it back on and it must be refused → Put the results in a table: how each attack was done, the request and response, and whether it was refused when it should have been → Save the tests as scripts in the repo and run them again before your next launch → When you're done, install Anthropic's official security-guidance plugin (/plugin install security-guidance@claude-plugins-official). After every turn of edits, and after each commit or push it makes, Opus reviews the changes again. Those reviews use extra usage

The default medium effort is enough. The impersonation group is the most expensive to get wrong, so raise that one to high.

Send this whole post to Claude Code. If you need to, add this line at the end 👇 "First check that the preview environment isn't connected to the production database. If it is, run everything locally only. Use test accounts and test data, with a separate set of accounts for each subagent. Don't touch real users, don't attack services I don't own, don't hammer any one endpoint, and only test payments in test mode. Only switch protections off temporarily in local code, and never change the database or dashboard settings. Keep test account passwords in .env, out of git, and mask tokens in the results table. Send one subagent per group to attack, recording without fixing. Run the impersonation subagent at high effort. Then fix everything yourself from their lists, rerun every test, and give me a before-and-after table. Delete the test accounts and data at the end. For installing the plugin or changing dashboard settings, give me the steps and I'll do it."