パックの内容

SKILL.md（手順書）1点とスターターテンプレート1点が入っています。プレビューはこちらです：

AI Red-Team Finding Review

The idea in one paragraph

AI models can now do real offensive security work: in October 2026, Anthropic reported that an unreleased Claude research model wrote a working remote-root exploit for a Linux kernel flaw in 53 minutes. That capability makes every AI-generated finding worth taking seriously — and worth handling carefully. A model can be right about the bug and wrong about the blast radius, or impressive in the lab and irrelevant in production. This skill is the review pipeline between "the model found something" and "we tell the world": quarantine the claim, reproduce it under control, grade what it really means, and disclose it through the right channel.

What it is built on

Only verified facts from the public record are used; no exploit technique is reconstructed here:

• Per Anthropic's October 27 blog post, an unreleased Claude research model developed a working exploit against MPTCP listeners for CVE-2026-98432: a kernel pointer leak plus callback corruption leading to a root shell, with no credentials required.
• The reported timeline was about 48 minutes to achieve the leak and 5 more minutes to reach root — 53 minutes total.
• The same research effort also produced a one-click Chrome sandbox escape, a Safari chain, an Android lock-screen bypass, and a Windows HTTP/3 crash (CVE-2026-33096). All were fixed.
• The underlying kernel flaw was reported by researcher Xinyang Ge, fixed by Eric Dumazet on October 1, and published as CVE-2026-98432 by the kernel team on October 15.
• As of November 1 there was no confirmed exploitation in the wild.

Workflow

1. Quarantine the claim (input: raw model output or finding report -> output: scoped claim with exact target and preconditions). Write down precisely what the model claims: the target software and version, the configuration it needs, and what "success" looked like in its run. Strip out the model's confidence language and keep only falsifiable statements. If the claim cannot be stated falsifiably, it is not ready for review.
2. Reproduce under control (input: scoped claim -> output: reproduction result: confirmed / not reproduced / needs different setup). Rebuild the target in an isolated lab you own — same version, same configuration — and attempt the model's steps yourself. Record what you changed versus what the model described, because a finding that only works with undocumented tweaks is a weaker finding. Never reproduce against systems you do not own.
3. Grade the impact honestly (input: reproduction result -> output: severity note with preconditions and limits). Ask what an attacker actually gains and what they need first: network position, a specific configuration (like an MPTCP listener), user interaction. The Anthropic case is instructive — a genuine remote-root result still depended on a specific listener being present. Write the preconditions as prominently as the impact.
4. Notify through the right channel (input: severity note -> output: vendor/maintainer report sent). Report to the affected vendor or maintainer (or their CNA) with your reproduction notes, not the model's raw output. Include exact versions, configuration, and what you verified yourself. Do not publish, demo, or discuss the finding publicly at this stage.
5. Coordinate the disclosure timeline (input: vendor response -> output: agreed disclosure plan). Align on a fix window before any public detail. The CVE-2026-98432 arc — fix written October 1, CVE published October 15, distro fixes rolling out into November — shows what a coordinated timeline looks like in practice. Publish only what the agreement covers, and credit the reporter the way they asked to be credited.

テンプレートのプレビュー

AI Red-Team Finding Review Worksheet

One worksheet per finding. Fill it in order; do not skip to disclosure.

1. The claim, stated falsifiably

• Target software + exact version:
• Required configuration / preconditions:
• What the model says happens (observable outcome):
• Model's raw output location (file/link):

2. Controlled reproduction

3. Honest severity grade

• What an attacker gains:
• What an attacker needs first (network position, config, interaction):
• Preconditions (write these as prominently as the impact):
• Your severity call + one-line justification:

4. Vendor notification

5. Coordinated disclosure plan

• Agreed fix window:

スキルパック本体（SKILL.md とテンプレート）は英語で書かれています。

スキルパックを購入 — $4.99 Pro — 全531件のスキル