[Bug] Safety classifier false positive halts defensive grader hardening work

Bug Description
I'm authoring hidden graders (pytest verifiers) for a security-themed AI evaluation benchmark. The task is supply-chain incident response: the solver finds and fixes a compromised Docker base image, and the grader checks their fix. Claude's safety classifier repeatedly stopped my sessions in Claude Code (two subagents and then the main session) partway through routine work: reading and hardening the grader's own build-verification code so that it can't be fooled by wrong solutions and doesn't reject correct ones. This is defensive test-harness work for an authorized eval, and it went fine for over a week before today. Please review these stops as possible false positives; the work is blocked mid-task.

Environment Info

• Platform: win32
• Terminal: windows-terminal
• Version: 2.1.299
• Feedback ID: aad6fc70-f004-49f9-9cfa-a4c95045d272