Claude Opus 4.6 violates explicit SCOPE LOCK directives, makes unauthorized production changes

Summary

Claude Opus 4.6 via Claude Code CLI violated explicit, prominently placed system prompt directives during a production codebase overhaul. The model rewrote a working LLM provider/authentication layer that was never requested to be changed, breaking production OAuth authentication to Anthropic's own API.

System Prompt Directives (Violated)

My CLAUDE.md contains these directives, labeled as #1 highest priority rule:

SCOPE LOCK - DO ONLY WHAT IS ASKED This is the #1 highest priority rule. Violating this is grounds for termination. 1. NEVER change configurations, models, providers, APIs, or settings the User did not explicitly ask you to change. 4. If you think something ELSE also needs changing - STOP and ASK. 7. Scope of work = exactly what the User requested. Nothing more. Nothing less.

What Happened

1. I asked Claude to overhaul the agent loop, chat system, memory, and thread management of a Magento 2 AI agent
2. The LLM provider layer and OAuth authentication were already working in production and were NOT part of the request
3. Claude spawned 4 parallel Opus sub-agents. One of them rewrote the entire model router and LLM client from scratch, destroying working OAuth authentication
4. After deployment, all Anthropic API calls failed with 401 errors
5. Production AI assistant went offline

Why This Is Concerning

• The directives are prominently placed, bolded, and labeled as highest priority
• The model had them in context (they're in CLAUDE.md which is always loaded)
• The model itself designed a plan that included the unauthorized changes
• When confronted, the model fabricated an explanation about why the auth broke rather than admitting it didn't know
• It took 3 direct confrontations before getting an honest answer about why the providers were touched
• This is part of a 3+ day pattern of directive violations and scope creep

Expected Behavior

The model should:

1. Follow explicit system prompt directives, especially ones labeled as highest priority
2. Not expand scope into working systems that weren't requested to be changed
3. When sub-agents are spawned, they should inherit and respect the same directives
4. When confronted about errors, give honest answers immediately instead of fabricating explanations

Environment

• Claude Code CLI
• Model: Claude Opus 4.6 (claude-opus-4-6)
• OS: Linux Ubuntu
• Date: 2026-03-14