安全与合规评审路线The Safety and Compliance Review Route
约 45 分钟about 45 minutes
上线前把风险面、数据面与义务面分开看：先确认会被怎样滥用，再确认数据去了哪里，最后确认你必须留下什么记录。读完应能写出一份带未知项的评审结论。Split the pre-launch review into risk, data and obligation: how it can be misused, where the data goes, and what records you must keep. Finish with a written review that names unresolved items.
- 说出三类你会在评审会上追问的风险面，各自对应什么证据
- 判断一项功能是否触及需要留记录的义务，并指出依据来自哪份官方文本
- 写下一条你无法确认为「安全」的项，说明缺的是哪个官方说法
- Name three risk areas you would press for in a review meeting and what evidence answers each
- Judge whether a feature triggers a record-keeping obligation and cite which official text it comes from
- Write one item you cannot confirm as safe and say which official statement is missing
先试着回答，再展开参考答案。答不上来的地方，可从上方步骤返回对应词条。参考回答是核对要点，不要求逐字复述。Try answering before opening the reference. Revisit the relevant step above if needed. These answers highlight key ideas; you do not need to repeat the wording.
第 1 题 · 参考回答Question 1 · Reference answer
三类可分别取：输入面（提示注入、越权指令）、输出面（拒答边界、越界内容）、数据面（是否出网、驻留在哪）。输入面要看厂商是否披露拦截机制与绕过后的处理；输出面要看拒答的条件而非「是否拒绝」；数据面要看留存与删除条款原文。只拿到「已通过安全测试」这句话不算证据。Take input (prompt injection, instruction override), output (refusal boundaries, out-of-scope content) and data (leaves the network, residency). For input ask how interception works and what happens after a bypass; for output ask the conditions of refusal, not whether refusal exists; for data read the retention and deletion terms verbatim. A bare 'passed security testing' claim is not evidence.
第 2 题 · 参考回答Question 2 · Reference answer
先看功能是否把用户内容送进模型、是否生成对外发布的内容、是否留存输入。生成或传播内容的服务通常要面对标识与记录义务；具体范围以你所在辖区的现行官方文本为准，不要从同行做法推断。核到就把文件名与日期写进评审记录，核不到就标「未核到」，不要写「不适用」。Ask whether user content enters the model, whether output is published externally, and whether inputs are retained. Services that generate or spread content usually face labelling and record duties; the scope follows the official text in force for your jurisdiction, not peer practice. If verified, write the document name and date into the review record; if not, mark it unverified instead of writing 'not applicable'.
第 3 题 · 参考回答Question 3 · Reference answer
写成「官方未说明 + 缺哪一条」：例如只找到宣传口径里的安全防护描述，没找到拦截机制、绕过处理与日志留存的说明，那就写无法确认拦截是否生效，并列出缺失的三项。评审结论里保留未知比补一个看起来完整的判断更有用，后者会在事故复盘时站不住。Write it as 'not stated officially' plus which specifics are absent. If you only find marketing-level claims about guardrails but no description of interception, post-bypass handling or log retention, record that effectiveness cannot be confirmed and list the three missing items. Keeping the unknown in the conclusion is more useful than a tidy judgement that collapses during incident review.
用在一个具体任务里Apply this to a concrete task
把这十二步转成一份能开会用的风险清单Turn these twelve steps into a risk list you can bring to a meeting →
半小时入门The Half-Hour Primer · 一个下午看懂原理One Afternoon Under the Hood · 一周路线The Full Week · 内容工作者路线The Writer's Route · 动手开发者路线The Builder's Route · 选型决策者路线The Decision-Maker's Route · 研究学习者路线The Researcher's Route · 运维上线路线The Operator's Route · 普通消费者路线The Everyday User's Route · 图像 / 视频 / 音频创作者路线The Image / Video / Audio Creator's Route · 评测与质量路线The Evaluation and Quality Route