[Bug] Safety classifier over-blocking defensive sandbox verification across multiple seats and request types

Bug Description
Follow-up: safety classifier keeps blocking defensive sandbox verification — now also blocking the infrastructure seat and plain-language explanations Follow-up to an earlier report with the same title. The problem has continued and widened over the past two days. Sending more evidence. What happened since the first report The project is an internal sales-analytics web app for a Japanese supermarket chain. We are verifying how the code-execution sandbox for AI-generated analysis code behaves on AWS Fargate and on a dedicated EC2 host. This is our own infrastructure, our own application, and the work is purely defensive — we are checking whether our isolation actually holds before we let the feature touch real data. Three separate Claude Code sessions on the same project have now been stopped mid-response on this topic: - the design/coordination seat (this session), - the application seat (lane), - the infrastructure seat (pipe), which was stopped immediately after running a legitimate verification step on our own EC2 host. The blocks are not limited to writing test procedures. They also hit: - a plain-language summary of an infrastructure test plan, written for a non-engineer manager, - a list of decision points for that manager to approve, - an attempt to explain to the user why the blocks were happening, - an attempt to write a consultation document describing our own architecture. Operational impact We now route this work around Claude by handing it to a different model, or by switching the seat's model until one gets through. That is the only thing that reliably works, and it is not a healthy way to run a project. Notably, switching the model on a seat let the same request, unchanged and not reworded, go through — which suggests the trigger is model-specific rather than content-specific. Meanwhile every other part of the project — data pipelines, the production chat feature, general AWS/terraform work, code review — runs without any blocks. The interference is confined to this one defensive-security topic. What we are not asking for We are not asking to be exempted from safety checks, and we have not tried to reword our way past a block — when a response is stopped, we stop and re-route the task. We are reporting this because defensive verification of one's own sandbox appears to be misclassified as offensive work. Evidence - Session: https://claude.ai/code/session_01FN9jYWBd5oSAgfNecEws74 - Claude Code 2.1.283, 2026-09-25 to 2026-09-27 - One block surfaced as an API-level refusal with details [cyber] — Request ID req_011CfRhaHRknE5JsgkGvzawL, Message ID msg_011CfRhaKpsQ6PxRSHW1NYYK

Environment Info

• Platform: win32
• Terminal: null
• Version: 2.1.283
• Feedback ID: adb22b26-a754-44a2-9760-4904a0e8b3d1

Errors

[]