← Safety

核验标记表示资料完整度，不是产品评分。下列未确认项可能影响选型；请按对应来源逐项确认。The confidence mark describes evidence completeness, not a product score. Unconfirmed items below may affect your choice; check them against the relevant sources.

English summaries and interface labels are available. Detailed source notes below are currently in Chinese.

模型与开放条件Model & openness

不开放选模型：检测引擎是厂商自有的，官方页面未披露底层模型档位，也没有自带模型的入口。它守的是你的 agent 与 LLM 链路——DLP API 收文本与文件，MCP Gateway 管「AI clients 与 remote MCP servers 之间」的流量，Inference Hooks 的官方 文档页标题点名 Cursor、Claude Code 与 VS Code Copilot。能否替换或自选检测模型：未取到官方声明，记未知。

No model picking is offered: the detection engine is the vendor's own, the official pages disclose no model tiers and give no bring-your-own-model entry point. What it guards is your agent and LLM chain — the DLP API takes text and files, the MCP Gateway sits between "your AI clients and remote MCP servers", and the Inference Hooks documentation pages name Cursor, Claude Code and VS Code Copilot. Whether the detection model can be replaced or chosen: no official statement captured, recorded as unknown.

运行位置Where it runs

托管交付：扫描走 api 侧的 /v3/scan 与 /v3/upload 系列端点，结果在控制台与 Violations 页回看（release notes 在册）。 端点侧确有本地代理——macOS/Windows 的 endpoint agent 在文档与更新日志里在册（AirDrop 与 Bluetooth 外发控制、 MDM 授权），但检测与判定按官方口径仍是服务侧能力。自托管或 air-gapped 形态：当日抓到的页面未列，记未知。

Hosted delivery: scanning goes through the /v3/scan and /v3/upload endpoints on the API side, with results reviewed in the console and on the Violations page (on record in the release notes). There is a local agent on the endpoint side — the macOS/Windows endpoint agent appears in the docs and release notes (AirDrop and Bluetooth exfiltration control, MDM-granted permissions) — but detection and verdicts remain a service-side capability per the official framing. A self-hosted or air-gapped form: absent from the pages captured, recorded as unknown.

本地文件Local files

你能交给它的东西（官方 FAQ 在册）：「plain text, structured and unstructured files, and even images」，格式点名 JSON、XML、CSV 等；端点面则在你的机器上读文件以便在读取时刻阻断。交给它的内容是否留存、留多久：DLP API 文档与 FAQ 均未声明，进 unknowns。图像与文件的能力限制见 supported file types 文档页（本批次未逐条读取）。

What you can hand it (official FAQ on record): "plain text, structured and unstructured files, and even images", with JSON, XML and CSV named; on the endpoint side the agent reads files on your machine in order to block at the moment of read. Whether submitted content is retained, and for how long: not declared in the DLP API docs or FAQ, moved to unknowns. File and image capability limits live on the supported-file-types page (not read item by item in this pass).

关机后的任务Tasks after shutdown

异步任务面清楚：文件扫描是 /v3/upload（POST 发起）→ /v3/upload/{fileId}（PATCH 传分块）→ /finish → /scan 的多步流程， 配套 webhooks 与异步通知文档页在册；事件在控制台聚合。这层跑在厂商侧，不存在「关机后续跑」的本机任务语义。

The asynchronous surface is explicit: file scanning is a multi-step flow — /v3/upload (POST to start), /v3/upload/{fileId} (PATCH chunks), /finish, then /scan — with webhook and asynchronous-notification pages on record, and events aggregated in the console. This layer runs vendor-side, so the "keeps running after shutdown" local-task semantic does not apply.

工具与扩展Tools & extensions

API 面（DLP 文档页在册）：请求字段 payload、policyUUIDs、detectionRules、redactionConfig、alertConfig，响应字段 findings、redactedPayload、message。扩展面：Nightfall 官方 MCP server 文档在册（让 AI 助手直接查安全数据）， Inference Hooks 接 IDE/agent，Premier 档另列 OpenTelemetry audit trail 与 MCP server discovery（pricing 页在册）。 MCP 传输方式与工具审批细节：本批次未逐条核验。

API surface (on record in the DLP docs): request fields payload, policyUUIDs, detectionRules, redactionConfig and alertConfig; response fields findings, redactedPayload and message. Extension surface: an official Nightfall MCP server documentation page (letting AI assistants query your security data directly), Inference Hooks into IDEs and agents, and Premier-tier OpenTelemetry audit trail plus MCP server discovery (on the pricing page). MCP transport details and tool-approval semantics: not verified item by item in this pass.

上下文与记忆Context & memory

它的「上下文」是策略与规则集，不是会话记忆：一次扫描由 policyUUIDs 与 detectionRules 决定按哪些规则跑。 可复核的历史面有明确边界——MCP Gateway 的审计记录写 user、server、tool、status 与 request payload， 而 payload「capped at 8 KiB」；超出部分的处置方式官方未说明，记未知。

Its "context" is policy and rule set, not conversation memory: one scan is defined by which rules run, via policyUUIDs and detectionRules. The revisable history has a stated boundary — MCP Gateway audit records carry user, server, tool, status and request payload, with payloads "capped at 8 KiB"; how anything beyond that is handled is not stated officially, recorded as unknown.

权限与限制Permissions & limits

逐条挂来源：SOC 2 Type 2 在 security 页在册；数据出境、留存期、是否用于训练这三问在本批次抓到的任何官方页面里 都没有逐条答案（developer-api 文档与 FAQ 未声明，/trust 404），因此一律进 unknowns，不做「不训练」这类推断。 档位与额度事实（pricing 页在册，不含价格）：Foundation Edition 覆盖「12 API-integrated SaaS apps」的 Data Detection & Response，端点代理「covers every SaaS app, desktop app, and AI app」，每用户许可含「2 endpoint devices」， 各档均含「150 GB of data-at-rest scanning」；Premier 另加 AI agent stack、MCP server discovery 与 OpenTelemetry audit trail。红线口径：官方在册的是检测类别与动作（block、mask/redact、removeFinding、cryptoConfig）， 本站不写检出率、漏检率，也不把它与自部署方案排名——选型看的是「数据能不能出你的边界」这一条约束。

Clause by clause with sources: SOC 2 Type 2 is on record at the security page; transfer path, retention period and training use of customer content have no itemised answer on any official page captured in this pass (neither the developer-api docs nor the FAQ declares it, and /trust returned 404), so all three go to unknowns with no inference such as "not used for training". Tier and quota facts (pricing page, no prices): Foundation Edition covers "Data Detection & Response across 12 API-integrated SaaS apps", the endpoint agent "covers every SaaS app, desktop app, and AI app", each user licence includes "2 endpoint devices", and every plan includes "150 GB of data-at-rest scanning"; Premier adds the AI agent stack, MCP server discovery and the OpenTelemetry audit trail. Red line: what is on record are detection categories and actions (block, mask/redact, removeFinding, cryptoConfig); this site publishes no detection or miss rate and does not rank Nightfall against self-deployed options — the selection question is whether your data may leave your boundary.

适合什么任务Best for

适合：SaaS、端点与 agent 三条面已经在用，要统一管控；需要在 MCP 与 IDE hooks 这一层加动作；愿意用托管 API 在发给 模型之前把 PII 与凭据处理掉；需要事件面与审计面而不是自己搭一套。不适合：硬约束「文本不出自己的边界」或要求可审计 的开源许可（那种约束下走自部署方案）；只需要凭据泄露监测与 incident 工作流的团队。

Fits: SaaS, endpoint and agent surfaces already in use and needing one control plane; needing real actions at the MCP and IDE-hook layer; happy to use a hosted API to strip PII and credentials before data reaches a model; wanting an events and audit surface rather than building one. Not a fit: a hard "text never leaves my boundary" constraint, or a requirement for an auditable open-source licence (that constraint points at self-deployed tooling); teams that only need leaked-credential detection with an incident workflow.

本页由This page is generated from the ai-handbook 数据层生成（CC BY 4.0）。data layer (CC BY 4.0). 方法论与坐标系定义见仓库内For the methodology and coordinate definitions, see METHODOLOGY.md.