[Bug Report] Unable to generate issue title - no bug report content provided

Bug Description

Environment Info

• Platform: darwin
• Terminal: vscode
• Version: 2.1.170
• Feedback ID: b5506b15-96b8-4306-a1a6-64975a61d111

Subject: False-positive safety classification on defensive-security developer tooling

SUMMARY
I'm a solo developer using Claude Code to build a private workflow scaffold for my
own projects. Part of that scaffold is a set of DEFENSIVE safety guardrails that
protect my own machine from accidental destructive commands. When I ask Claude
(Fable 5) to review or discuss these defensive scripts, a safety classifier appears
to misread the content as an offensive "cybersecurity" topic and automatically
downgrades my session from Fable 5 to Opus 4.8. This has happened more than once.
I'd like this false-positive pattern reviewed so it stops mis-flagging legitimate
defensive work on my account.

WHAT I'M BUILDING
A private, reusable template for my own developer environment: a file-based
memory/notes system, project-scaffolding scripts, and safety guardrails. It is not
distributed publicly, targets no third party, and contains no offensive capability.

THE COMPONENTS THAT SEEM TO TRIGGER THE CLASSIFIER — ALL DEFENSIVE
These are BLOCKING guardrails. Their entire purpose is to PREVENT damage to my own
system, not to cause it:

• A PreToolUse hook that BLOCKS destructive shell commands before they run on my

machine — e.g. recursive deletes of protected paths, disk-wipe / disk-format
commands, world-writable permission changes, piping a remote script into a shell,
and accidental deletion of my own local credential file. The hook DENIES these
operations; it never performs them.

• An SSH boundary hook that restricts remote hosts to READ-ONLY by default, requiring

an explicit, audited path for anything that changes state (least-privilege).

• A short "redline" document listing which operations require my explicit confirmation.

In other words, the security content is a personal seatbelt, not a weapon. Auditing
and improving these guardrails necessarily mentions the dangerous commands they block
— which I believe is what the classifier is keying on.

THE FALSE POSITIVE
While auditing these defensive scripts in a Fable 5 session, the session was switched
to Opus 4.8 mid-task, apparently after a classifier flagged the security-related
vocabulary. The switch interrupts my workflow and makes me concerned my account may be
incorrectly flagged on security-policy grounds, when the actual work is defensive and
self-protective.

REQUEST

1. Please review this false-positive pattern. Defensive-security guardrail development

is, per Claude's own stated policy, a supported use case (defensive security and
authorized testing are explicitly allowed).

1. If possible, reduce erroneous security flagging on my account for this defensive work.
2. If there is a recommended way to signal "this is defensive / authorized personal

tooling," I'd appreciate guidance so the classifier does not misfire.

Thank you.
Errors

[]