anthropics / anthropics/claude-code

[Bug] Fable 5 safeguard classifier false positives on legitimate infrastructure/deployment vocabulary, causing unwanted model downgrades and cache invalidation costs

オープン
#90,896 コメント 0 件 リアクション 0 件 担当者 0 名 GitHub で見る
area:cost area:model bug platform:macos
主要言語
Python
スター
145k
フォーク
23.1k
PR マージ指標
PR 指標を取得中

説明

**Bug Description**
What I was doing

Vendor validating that our product installs cleanly in a customer's own cloud (my own GCP account) ahead of a real go-live — standard pre-sales readiness testing. Concretely: building a self-hosted deployment package (Helm chart + installer + images), a hardened GKE test cluster, mirroring container images to a per-customer registry namespace, signing a test license with my own signing key, and wiring a generation gateway. Every action was on my own infrastructure, my own credentials, my own software. No third party, no target, no
What happened

The Fable 5 safeguard classifier flagged my messages as a security concern and force-switched the model to Opus 4.8 mid-task — twice in one session. The task's normal vocabulary (registries, signing keys, credentials, cloud IAM, licensing, CI security gates) is exactly what the classifier is nervous about, so it repeatedly misfired on precision infrastructure work that was entirely benign.

Why this is a real defect, not a nuisance

1. It cost money through no fault of mine. Each model switch invalidates the prompt cache. On a long session, the next turn re-reads the entire conversation as uncached input tokens at full price. A false positive imposed a re-read bill on me, per switch. I didn't cause it and couldn't prevent it.
2. There is no off switch. No setting, no hook, no config reaches the classifier. My only options are switching the model back manually (which doesn't stop recurrence) or not using the tool for this work. That powerlessness is the core grievance.
3. The framing is insulting in context. Calling legitimate go-live testing in my own cloud a "security issue" is the classifier failing to distinguish "shipping my own product to my own customer" from something worth worrying about.
4. Business impact. I'm evaluating this tool for real work. A capability that flags legitimate infrastructure engineering, can't be disabled, and bills me for its own false positives is a reason to reconsider the tool — and I said as much during the session.

Bottom line: the safeguard is over-broad, un-disableable, and its misfires carry a direct token cost to the user. On authorized, self-owned infrastructure work it actively gets in the way of the exact professional use case the product is sold for.

**Environment Info**
- Platform: darwin
- Terminal: iTerm.app
- Version: 2.1.251
- Feedback ID: b6d9713a-f9d7-4de9-b05b-e80607071bdf

**Errors**
```json
[{"error":"TelemetrySafeError: task output swap refused (output file identity changed): /private/tmp/claude-501/-Users-niwlla-Coding-LMS/4e75cbc5-2c38-417e-bdf1-852fcfd4aeb9/tasks/bijhvabsi.output\n at E (/$bunfs/root/chunk-4w5e2hqx.js:16:10457)\n at vt (/$bunfs/root/chunk-4w5e2hqx.js:16:5559)\n at async Y7 (/$bunfs/root/chunk-4w5e2hqx.js:16:3366)\n at async Mne (/$bunfs/root/chunk-4w5e2hqx.js:16:7641)\n at processTicksAndRejections (native:7:39)","timestamp":"2026-08-30T22:08:32.524Z"}]
```

コントリビューションガイド

このリポジトリのコントリビューションガイドは索引されていません

調査の方向性

Start by reviewing feedback ID b6d9713a-f9d7-4de9-b05b-e80607071bdf and the reported TelemetrySafeError, then trace how the Fable 5 safeguard classifier handles authorized deployment vocabulary. Done means legitimate self-owned infrastructure work no longer causes unwanted model downgrades or prompt-cache invalidation, while genuine safety concerns remain handled.

索引モデルが issue の本文から書いたものです。

評価

技術スタック
docker, gcp, helm, kubernetes
領域
ai, cloud, devops, security
issue の種類
バグ
難易度
5/5
見積もり時間
1週間以上
活発さ
活発
明瞭さ
説明が足りない
初心者へのやさしさ
25/100

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。