Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
AIの権限設定に悩んでいる話
Search
Sponsored
·
SiteGround - Reliable hosting with speed, security, and support you can count on.
→
Shuma
March 19, 2026
Programming
48
0
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
AIの権限設定に悩んでいる話
2026/3/19(木) 第四回 唐揚げ会 発表LT
Shuma
March 19, 2026
More Decks by Shuma
See All by Shuma
インフラ深掘りLT
shubox
0
47
飲食店長から_SREになった話
shubox
0
35
Ansible で Vector を導入し Slack 通知とログレベル色分けまでした話
shubox
0
55
阿部寛のホームページをSRE観点で改善出来るか考えてみた。
shubox
0
140
一日の終わりに、晩酌しながら眺めたいシステムログの世界
shubox
0
130
プロダクトがクローズした話
shubox
0
170
今も熱いもの!魂を揺さぶる戦士の儀式:マオリ族のハカ
shubox
0
310
信頼性工学とは? ~カツオを題材に~
shubox
0
130
バージョン管理と人生設計の類似点ver2 -リポジトリから学ぶ生きるヒント-
shubox
0
310
Other Decks in Programming
See All in Programming
Hatena Engineer Seminar #37「言語モデルの活用に関する研究」
slashnephy
0
460
【やさしく解説 設計編 #0】DDDのコード、読めるのに分からない人へ
panda728
PRO
1
170
Snowflake Summitでの新機能 CoCo / CoWork / snowflake-summit-2026-overall-what-new-coco
tatsuhiro
1
210
任せる範囲はこう広がった / How the Scope of AI Delegation Has Expanded
nrslib
1
220
コンテキストの使い捨てをやめる — ビジネスルール駆動開発と miko —
ioki
0
260
代数的データ型って何が嬉しいの? #frontend_phpcon_do
kajitack
8
3.9k
act1-costs.pdf
sumedhbala
0
150
Hunting Vulnerabilities in Symfony with LLMs
vinceamstoutz
0
570
Skillsは効率化、Agentsは"自分の拡張"——Builder時代のエージェント編成(CC Night 2026)
wemra
1
190
フロントエンドとバックエンドで「1文字」を揃えよう
youkidearitai
PRO
0
780
The Bowling Game- From Imperative to Functional Programming - Part 1
philipschwarz
PRO
0
210
OS アップデート対応の取り組み方がもっと共有されてほしい
andpad
0
110
Featured
See All Featured
Exploring the Power of Turbo Streams & Action Cable | RailsConf2023
kevinliebholz
37
6.5k
Automating Front-end Workflow
addyosmani
1370
210k
I Don’t Have Time: Getting Over the Fear to Launch Your Podcast
jcasabona
34
2.8k
Connecting the Dots Between Site Speed, User Experience & Your Business [WebExpo 2025]
tammyeverts
11
960
Collaborative Software Design: How to facilitate domain modelling decisions
baasie
1
260
Pawsitive SEO: Lessons from My Dog (and Many Mistakes) on Thriving as a Consultant in the Age of AI
davidcarrasco
0
180
Jamie Indigo - Trashchat’s Guide to Black Boxes: Technical SEO Tactics for LLMs
techseoconnect
PRO
0
250
The World Runs on Bad Software
bkeepers
PRO
72
12k
The Organizational Zoo: Understanding Human Behavior Agility Through Metaphoric Constructive Conversations (based on the works of Arthur Shelley, Ph.D)
kimpetersen
PRO
0
380
10 Git Anti Patterns You Should be Aware of
lemiorhan
PRO
659
62k
Designing for Performance
lara
611
70k
jQuery: Nuts, Bolts and Bling
dougneiner
66
8.5k
Transcript
1 2 3 4 5 6 >_ AIは愛を知らないが、 エンジニアは AIを知っている
〜AIの権限設定に悩んでいる話〜 NotebookLM
自己紹介 名前: Shuma 職種: SRE / インフラエンジニア 興味: @ShuShuShuBOOOx 最近うれしかったこと
IaC, 障害対応, 監視, パフォーマンス改善 大きな負荷対策の監視をしていて 同じ会社の人とかなりのトラフィックをさば いたこと。
1 2 3 4 5 6 7 8 9 10
11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 「AIツール、業務で使ってますか?」 [ ChatGPT ] [ GitHub Copilot ] [ Claude Code ] [ Cursor ] 「権限設定してますか?」
$ Claude Code + Terraform で本番RDSが全消えした 月$5-10ドルの節約 → 2.5年分のデータが消えた (※24時間後に復旧)
> terraform destroy を承認したのは人間だった > 復旧対応でBusiness Supportにアップグレードしコスト10%増 → ソフトウェアとインフラ、両方が無防備だった (DataTalks.Club / Alexey Grigorev氏の事例)
None
settings.json 1 2 3 4 5 6 7 8 9
10 11 12 13 14 15 16 17 18 自分の実話①: deny ルールを設定した { "permissions": { "deny": [ "Bash(aws *)", "Bash(terraform *)", "Bash(kubectl *)" ] } } ✅ → 破壊的操作を全部止めた。方向性は正しい
>_ 自分の実話②:でも問題が出た 自分の実話②:でも、大きな問題が出た 1 aws * をdenyすると自分の作業まで邪魔してしまう 2 人間とClaudeが同じ .aws/credentials
を使っている 3 shared credentialsを使っているため、 Claudeの権限を絞ると人間の権限も絞れてしまう → じゃあどうする?
>_ 社内で出たさまざまな案 社内で出たさまざまな案(どれも帯に短し) 案A: credentialsをreadonly専用に 人間が不便 ✕ 案B: AI専用IAMユーザー作成 管理コスト増・攻撃面拡大
✕ 案C: 環境変数で分離 シェル履歴に平文で残るリスク ✕ 案D: SSO (IAM Identity Center) で分ける 初期設定・運用コストが高すぎる ✕ 案E: CLAUDE.md に禁止事項を書く 強制力なし ✕ → 完璧な単一解はなかった
> _ 現状の落としどころ:多層防御(第 1層) 1 { 2 "permissions": { 3
"deny": [ 4 "Bash(aws * delete*)", 5 "Bash(terraform destroy *)", 6 "Bash(terraform destroy)", 7 "Bash(terraform apply *)", 8 "Bash(terraform apply)", 9 "Bash(terraform state rm *)", 1 0 "Bash(terraform state rm)", 1 1 "Bash(aws iam *)", 1 2 "Bash(aws rds delete*)" 1 3 ] 1 4 } 1 5 } 第1層 Claude Code側の設定 deny ルール / Hooksでブロック → AIに「やらせない」 describe や plan は通す。 ⚠ ※aws s3 rmなどはすり抜ける可能性ありと注記
> _ 現状の落としどころ:多層防御(第 2層) 第1層(Claude Code側の設定) deny / Hooks →
AIに「やらせない」 第2層(インフラ側の物理的保護) 万が一指示しても「インフラ側で弾く」設定 IAM 最小権限 RDS 削除保護 Terraform prevent_destroy リモート State管理 AIに「やらせない」設定とインフラ側で「弾く」設定。この両輪が必要。
1 2 3 4 5 6 > まとめ:AIセキュリティの現在地 Q1: AIに本番環境を触らせるべきか?
Q2: denyルールはどこまで設定すべきか? Q3: インフラ側の保護は十分か? 結局、完璧な「正解」はない。 技術とリスクのバランスを見極め、 多層防御で対応していく必要がある。
1 2 3 4 5 6 7 8 9 10
11 12 13 › >_ Resources Claude Code official docs → docs.anthropic.com/en/docs/claude-code/settings Alexey Grigorev's Substack → alexeyondata.substack.com/p/how-i-dropped-our-production-database AWS IAM least privilege → docs.aws.amazon.com/IAM/latest/UserGuide/best-practices.html