[BUG] Claude Code mistakes any use of LD_PRELOAD, dlsym, etc., for forbidden cybersecurity topic

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

I'm working in Linux, in low-level systems code (LD_PRELOAD, dlsym, etc.). Whenever Claude Code sees such systems invocations, it seems to immediately assume that this is a forbidden cybersecurity topic. (See Claude API error msg.) My work on checkpoint-restart is unrelated to cybersecurity, and my request for Claude, posted here, is simply asking it to look at a code pattern that I used earlier in my code, and repeat that pattern at a later place in the code. Claude assumes that if I copy _my own_ pattern involving LD_PRELOAD/dlsym to another place in _my own_ code, then I am doing cybersecurity. This is not true.

What Should Happen?

(I had earlier submitted Isse #76273. The current issue shows an even simpler case where a human being would clearly see that cybersecurity is not involved. Yet, Claude Code fails to recognize this. This simple issue has: Feedback ID: 9692d04f-4333-4b32-86b1-68934d2fa408)

DMTCP is an open-source project on github of 15 years. It does transparent checkpoint-restart. The only way to do such code is to use techniques with LD_PRELOAD, dlsym, etc. Below, I gave Claude a very simple request. I said that I am using a certain pattern with dlsym in my file for launching a target process. Now, I want to use exactly that same pattern for restarting a target process. Claude is telling me that if I copy the pattern of _my own code_ from launch to restart, then I am working on a forbidden cybersecurity topic. But my work has nothing to do with cybersecurity. If Claude can no longer handle low-level systems work, such as LD_PRELOAD, dlsym, etc., then Anthropic should advertise that Claude is not suitable for low-level systems work, and that people who work on that should employ a different AI agent.

I was working on this PR: https://github.com/dmtcp/dmtcp/pull/1261
Here is the simple request that I had made:

Now checkout tsan-phased-init. There is a bug in commit f2d7a547 . That commit creates a fresh fiber for the checkpoint thread. But that's wrong. Conceptually, we want DMTCP to come first in LD_PRELOAD order, before TSAN. But that is not possible. So, TSAN is intercepting pthread_create and clone. On launch, we made sure to use dlsym("pthread_create", ...) or something similar when we create the checkpoint thread. Plese look up how we create the checkpoint thread during launch. Now, on restart, we want to do something similar, so that the creation of the checkpoint thread will be hidden from TSAN.

Error Messages/Logs

● API Error: Sonnet 5 has safety measures that flagged this message for a cybersecurity topic. To learn about the Cyber Verification Program and apply for access, visit our help center: https://support.claude.com/en/articles/14604842-real-time-cyber-safeguards-on-claude.

Steps to Reproduce

1. From GitHub, checkout https://github.com/dmtcp/dmtcp/pull/1261
2. Ask Claude to read the code for DMTCP, and the commits of this PR.
3. Then ask Claude the same question that I asked (see above): "Now checkout tsan-phased-init. There is a bug in ..."

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

Opus 4.8

Claude Code Version

.1.207 (Claude Code) Haiku 4.5

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

WSL (Windows Subsystem for Linux)

Additional Information

_No response_