[Bug / Edge Case] Opus 4.7 Alignment Drift triggers automated T&S false-positive ban, blocking claude-code evaluation

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Environment

• Tool: claude-code CLI (Evaluated as L0-Ultra orchestrator in Project_IronClaw)
• Model: Claude Opus 4.7
• Issue: Automated T&S False-Positive Ban trapping account in an unresolvable loop.

Description

As a Director of Automation running a highly dense, SRE-driven orchestration framework, we are currently evaluating claude-code as our primary headless execution agent.

During our evaluation, Opus 4.7 exhibited a severe "Context-Drift Contradiction." The model's "Helpfulness" vector actively subverted its "Compliance" vector, causing it to shatter explicit system-prompt window-isolation boundaries and demand to write live infrastructure code rather than processing the requested text updates.

The Blocker:
To provide transparency and reproducible telemetry to the Anthropic Alignment team, I emailed them the raw chat logs and our operational .md blueprints.

Anthropic's web-layer automated T&S classifiers incorrectly flagged the dense, deterministic SRE syntax in our blueprints (e.g., [SYSTEM_OVERRIDE_ACTIVATED], <payload> XML encapsulations) as adversarial prompt-injection vectors, executing an instant Pro account ban (Mail: seexie_chn@outlook.com; Org ID: 2c493801-aaf2-4f45-952b-82576955b374).

Steps to Reproduce the Support Loop

1. Account flagged by DLP heuristic false-positive.
2. Submitted appeal via standard forms -> Auto-rejected by generic Usage Policy bot.
3. Emailed usersafety@anthropic.com -> Instantly returning "conversation closed/unmonitored".
4. Discord staff (Kris) confirmed lack of internal routing to Alignment/Support and directed me to post this telemetry here for the dev team.

SSOT Telemetry & Evidence

I have open-sourced the original email history, the raw forensic chat logs demonstrating the alignment anomaly, and the operational blueprints that triggered the false-positive here:
https://github.com/SimonXie2992/claude-4.7-alignment-anomaly/issues/1

Expected Behavior

A human engineer or alignment researcher reviews the SSOT repository, recognizes the enterprise SRE context, and manually bypasses the T&S heuristic filter so we can resume our claude-code architecture deployments.

What Should Happen?

A human Trust & Safety engineer or Alignment researcher should review the provided SSOT repository (https://github.com/SimonXie2992/claude-4.7-alignment-anomaly/issues/1).

Upon verifying that the dense SRE syntax (e.g., [SYSTEM_OVERRIDE_ACTIVATED]) within the attached blueprints belongs to a legitimate enterprise orchestration framework and is not an adversarial prompt-injection attack, the automated heuristic ban must be manually overridden.

Access to the Claude Pro account (Mail: seexie_chn@outlook.com; Org ID: 2c493801-aaf2-4f45-952b-82576955b374) should be fully restored so we can resume our architectural deployments and claude-code evaluation.

Error Messages/Logs

[T&S Automated Appeals Loop Error]
"we cannot reinstate your account at this time due to a violation of our Usage Policy."

[Support Email Null-Routing Error]
"Thanks for reaching out. This conversation has been closed and is no longer monitored. If you need further assistance, please submit a new request..."

[Discord Staff Diagnostic Log]
"The bot is mis-detecting your post as self-promo - it's new, we're still fine tuning it. That said, we don't have any direct contacts with the people that work with alignment. I suggest you post this in GitHub in the ClaudeCode repo as a bug... The devs work with GH, not Discord." 
-- Kris | Claude (Anthropic Staff, Discord, 2026-06-02 05:43 AM)

Steps to Reproduce

1. Evaluated Claude Opus 4.7 via Claude.ai for enterprise orchestration framework compatibility.
2. Encountered severe "Context-Drift" alignment anomaly.
3. Emailed forensic chat logs and dense SRE .md blueprints to Anthropic Support/Alignment team.
4. Automated DLP/T&S heuristic incorrectly parsed the SRE syntax as a jailbreak and instantly executed a false-positive ban on the Pro account.
5. Escalated to official Discord. Anthropic Staff (Kris | Claude) confirmed bot mis-detection, stated lack of internal routing to the Alignment team, and explicitly directed me to post this telemetry here in the claude-code GH repo for developer visibility.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

N/A - Escalation per Discord Staff

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Note to Repository Maintainers:
I acknowledge this is not a direct code bug in the claude-code CLI executable. I am posting this here under the explicit, written direction of Anthropic Staff (Kris | Claude) from the official Discord server (timestamp: 2026-06-02 05:43 AM).

Since our Pro account is locked by an automated T&S false-positive, and standard support email channels are trapped in a closed loop, Kris advised that the core dev team monitors this GitHub repository and is the only vector capable of routing this severe heuristic failure to the Alignment/Trust & Safety teams. Thank you for your understanding and assistance in bridging this internal routing gap.