[Bug] Safety classifier incorrectly flags authorized homelab/personal network administration and silently downgrades model tier

Bug Description
Summary: The safety/"cyber" classifier repeatedly interrupts routine, clearly-authorized homelab and personal-network work, and each interruption appears to downgrade the model mid-task (Opus 5.5 → a lower tier). Over one session it fired ~5 times on work that was unambiguously my own infrastructure or access I was plainly granted.

The contexts it flagged — none remotely abusive:

1. Connecting to a GPU server that a friend explicitly shared to me over Tailscale and told me to use freely. Standard remote login to a machine I was given access to.
2. Administering my own Raspberry Pi DNS server on my own home LAN — routine setup and troubleshooting of hardware I own, on my own network.
3. Standard multi-machine work across my own fleet.

Two problems:

• The classifier isn't weighing ownership/authorization context that is present in the conversation. Whether the target is the user's own device, or access a named owner explicitly granted, is stated plainly right there every time, and it doesn't seem to factor in before the flag fires. "Admin my own network / use a box a friend shared with me" is the most normal thing in the world for this product's audience.

• The mid-task model downgrade is the bigger operational hit. Each flag seems to knock the session down a model tier, so my assistant gets abruptly less capable in the middle of real work, with no notice and no way to understand why from the user side. I have to manually reset the model each time.

Impact: For anyone using Claude Code on a homelab, a self-hosted stack, or a Tailscale mesh — a large and growing slice of the power-user base — this makes the tool feel like it's treating normal sysadmin as hostile, and silently degrades quality when it does.

Ask: Weigh clearly-stated authorization and device-ownership context before the cyber classifier fires on local/mesh networking; and decouple a classifier flag from an automatic, silent model downgrade — at minimum surface it, ideally don't degrade capability on false positives at all.

Environment Info

• Platform: linux
• Terminal: windows-terminal
• Version: 2.1.292
• Feedback ID: 31c84d2c-4900-44cd-8829-95c85c29582d