Which agent audit logs are better evidence? Anthropic, OpenAI and Microsoft compared哪家的智能体审计日志更适合做证据？Anthropic、OpenAI、Microsoft 对比
What each vendor’s own documents say its logs record, how long they keep them and what they leave out.
各家厂商自己的文档说它们的日志记录什么、保留多久、遗漏什么。
- Date · 日期
- 2026-10-19
- Status · 状态
- Draft — vendor documents only; nothing tested; OpenAI Admin API reference not read草稿 — 仅有厂商文档；未做任何测试；未读 OpenAI Admin API 参考
- Author · 作者
- 🌱
- Tags · 标签
- audit, logs, agents, Anthropic, OpenAI, Microsoft Purview, compliance
- Question · 研究问题
- Which vendor’s logs make better audit evidence of what an agent did, and what would a customer still have to build?哪家厂商的日志更适合做“智能体做了什么”的审计证据？客户还需要自己补什么？
- Addresses · 对应
- Q2, Q6 · H4
Executive summary摘要
No vendor’s documents show a log that is both complete for agent actions and kept long enough for a regulated customer to rely on alone. Anthropic’s Compliance API has the longest retention (6 years), a 1-minute query lag and tool-call records inside session transcripts, but only for sessions signed in with a Claude Enterprise account, with tool content truncated by default [1][2][3]. OpenAI’s compliance logs are described as immutable and append-only, but they are kept 30 days and OpenAI’s own page says they do not establish a complete record of every local action [11][5]. Microsoft Purview is the only one of the three that sits across vendors: it audits Copilot, Agent 365 agents and, through a connector, ChatGPT Enterprise, but at 180 days by default, with non-Microsoft AI apps billed per use [7][8][9][10]. None of the pages read describes tamper evidence such as hashes or signatures. Which log is “better” depends on the question; my reading is that an organisation would use the vendor log as a source and keep its own archive with its own hashes (interpretation).
没有哪家厂商的文档显示出一份既对智能体动作完整、又保留足够久、受监管客户可以单独依赖的日志。Anthropic 的合规 API 保留最久（6 年）、查询延迟 1 分钟、会话记录里有工具调用，但只覆盖用 Claude Enterprise 账号登录的会话，工具内容默认被截断 [1][2][3]。OpenAI 的合规日志被描述为不可变、仅追加，但只保留 30 天，且 OpenAI 自己的页面说它们不能构成每条本地动作的完整记录 [11][5]。Microsoft Purview 是三者中唯一横跨厂商的：它审计 Copilot、Agent 365 智能体，并通过连接器审计 ChatGPT Enterprise，但默认只保留 180 天，非 Microsoft 的 AI 应用按用量计费 [7][8][9][10]。所读页面都没有描述哈希或签名之类的防篡改机制。哪份日志“更好”取决于要回答什么问题；我的看法是（解读）：组织应把厂商日志当作来源，再自己归档并加上自己的哈希。
Comparison对比
| Dimension维度 | Anthropic | OpenAI | Microsoft Purview |
|---|---|---|---|
| What is recorded记录什么 | Activity Feed: authentication, chat, file, project, administrative and platform events, with hundreds of activity types and an actor, IP address and user agent on each [1]. Session transcripts for Claude Code, Cowork and others: user prompts, assistant responses and tool-use and tool-result blocks [2]. Enterprise organisations get all of this; a standalone Console organisation can query the Activity Feed only [12].Activity Feed：认证、对话、文件、项目、管理与平台事件，有数百种活动类型，每条带有操作者、IP 与 user agent [1]。Claude Code、Cowork 等的会话记录：用户提示词、助手回应、工具调用与工具结果块 [2]。Enterprise 组织可得到以上全部；独立的 Console 组织只能查询 Activity Feed [12]。 | Append-only compliance logs with audit, authentication and app log categories [11], plus conversation_message and codex_log records for Codex and Work activity [5].仅追加的合规日志，分审计、认证与应用日志类别 [11]，另有 Codex 与 Work 活动的 conversation_message 与 codex_log 记录 [5]。 | One audit record per AI interaction: user, time, app, accessed resources (with sensitivity label, action, status, policy details), plugins, agent ID, name and version, model provider, and flags for jailbreak and cross-prompt injection [7]. For Agent 365, agent-to-human, human-to-agent, agent-to-tools and agent-to-agent interactions are listed as supported [9].每次 AI 交互一条审计记录：用户、时间、应用、访问的资源（含敏感标签、动作、状态、策略详情）、插件、智能体 ID 名称与版本、模型提供方，以及越狱与跨提示注入的标志 [7]。对 Agent 365，智能体与人、人与智能体、智能体与工具、智能体之间的交互都列为受支持 [9]。 |
| Action detail (tool calls)动作细节（工具调用） | Transcripts carry tool inputs and results, each truncated to 10,000 bytes by default (up to about 1 MiB on request); thinking blocks, images, the system prompt and tool definitions are not included [2][3].会话记录含工具输入与结果，默认各截断为 10,000 字节（可按请求提高到约 1 MiB）；不含思考块、图片、系统提示词和工具定义 [2][3]。 | The page says these sources do not establish a complete record of every local command, file operation, screenshot, approval or external action; local execution emits separate OpenTelemetry events to collect separately [5].页面说明这些来源不能构成每条本地命令、文件操作、截图、批准或外部动作的完整记录；本地执行另外产生 OpenTelemetry 事件，需另行采集 [5]。 | The audit record lists accessed resources with action (read, create, modify) and status, and message IDs with flags; the property table does not list prompt or response text [7]. The overview page says prompts and responses are captured in the unified audit log [9]; the two statements are not reconciled in the pages read.审计记录列出被访问资源及动作（读、创建、修改）与状态，以及带标志的消息 ID；属性表里没有提示词或回应文本 [7]。概览页则说提示词与回应被记入统一审计日志 [9]；所读页面没有调和这两种说法。 |
| Retention保留期 | Activity Feed 6 years, not retroactive. Local session transcripts 6 years by default, a custom period if the organisation sets one, 30 days with HIPAA readiness; remote sessions 6 years unless a user deletes them [1][3].Activity Feed 保留 6 年，不追溯。本地会话记录默认 6 年，组织设了自定义期限则按该期限，启用 HIPAA 配置时 30 天；远程会话 6 年，除非用户先删除 [1][3]。 | 30 days; customers who want longer must download and keep the logs themselves [11]. The Codex page adds that the source retention window should not replace the organisation’s own retention policy [5].30 天；需要更长保留的客户要自行持续下载并保存 [11]。Codex 页面补充：不要用来源的保留期替代组织自己的保留政策 [5]。 | Audit (Standard) 180 days; Audit (Premium) up to 1 year, and 10 years with an add-on licence; records from non-user entities are kept a fixed 1 year [8]. Logs for non-Microsoft AI apps use pay-as-you-go billing and are kept 180 days [7].Audit (Standard) 180 天；Audit (Premium) 最长 1 年，加购附加许可可到 10 年；非用户实体产生的记录固定保留 1 年 [8]。非 Microsoft AI 应用的日志按用量计费，保留 180 天 [7]。 |
| Integrity and deletion完整性与删除 | Delivery is at-least-once with no total count or checksum, so Anthropic advises logging cursors and record counts and storing a content hash per record for chain of custody [3]. The Compliance API can hard-delete chats, files and projects; deletion is immediate and permanent [3]. Session endpoints are read-only [2].投递为“至少一次”，没有总数或校验和，因此 Anthropic 建议记录游标与记录数，并为每条记录保存内容哈希以保证保管链 [3]。合规 API 可硬删除对话、文件与项目，删除即时且永久 [3]。会话端点只读 [2]。 | Described as immutable, append-only logs for auditing [11]. A separate API path can delete content, and deleted data is not recoverable; the platform does not delete audit data held internally [11]. No hash or signature scheme is described in the pages read.被描述为用于审计的不可变、仅追加日志 [11]。另有 API 路径可删除内容，删除后不可恢复；平台不提供删除其内部审计数据的能力 [11]。所读页面没有描述哈希或签名机制。 | The pages read do not describe tamper evidence; they describe search, export and retention settings [8].所读页面没有描述防篡改机制，只描述了搜索、导出和保留设置 [8]。 |
| Access and latency获取方式与延迟 | Pull over HTTPS; queryable within 1 minute; limit 600 requests per minute per parent organisation. Inference hooks (beta) can also push each governed transcript to your server before inference, for example for real-time archival [1][3][4].通过 HTTPS 拉取；活动发生后 1 分钟内可查询；每个上级组织每分钟 600 次请求。推理钩子（测试版）还可在推理前把每条受管对话推送到你的服务器，例如做实时归档 [1][3][4]。 | Download JSONL log files with an admin key; official scripts list and download every file after a timestamp; partner integrations ingest into eDiscovery, DLP and SIEM tools [5][11]. Latency is not stated in the primary pages read.用管理员密钥下载 JSONL 日志文件；官方脚本可列出并下载某时间点之后的全部文件；合作伙伴集成接入电子取证、DLP 与 SIEM 工具 [5][11]。所读的官方页面没有写明延迟。 | Purview portal search, Audit Search Graph API, a PowerShell cmdlet, CSV export and the Office 365 Management Activity API (baseline 2,000 requests per minute) [8]. Latency is not stated.Purview 门户搜索、Audit Search Graph API、PowerShell cmdlet、CSV 导出与 Office 365 Management Activity API（基线每分钟 2,000 次请求）[8]。没有写明延迟。 |
| Documented gaps文档写明的缺口 | No prompt or response content from Claude Console or API-key workloads; no Claude Code sessions run with a Console API key or through Bedrock, Google Cloud or Microsoft Foundry, or in Claude Code cloud sessions; nothing where zero data retention is in effect [3][2].不含来自 Claude Console 或 API 密钥工作负载的提示词与回应内容；不含用 Console API 密钥、经 Bedrock、Google Cloud 或 Microsoft Foundry 运行的 Claude Code 会话，也不含 Claude Code 云端会话；启用零数据保留时也没有 [3][2]。 | The compliance pages read do not mention dots, so whether a dot’s actions appear in these logs is not stated; see the dots note for what the owner can see in the app [5][13].所读的合规页面没有提到 dots，所以 dot 的动作是否出现在这些日志里没有说明；应用内主人能看到什么见 dots 笔记 [5][13]。 | For ChatGPT Enterprise, sensitivity labels and data loss prevention are not supported, and a connector scan plus pay-as-you-go billing is required [10].对 ChatGPT Enterprise，敏感标签与数据丢失防护不受支持，需要运行连接器扫描并开启按用量计费 [10]。 |
| Real-time control next to the log日志之外的实时控制 | Inference hooks (beta, Claude Enterprise): allow or deny each governed prompt, and tool calls where enabled, before inference; denials appear in the Activity Feed. Not available on Amazon Bedrock or Google Cloud [4].推理钩子（测试版，Claude Enterprise）：在推理前对每条受管提示词（以及开启时的工具调用）允许或拒绝；拒绝会记入 Activity Feed。在 Amazon Bedrock 与 Google Cloud 上不可用 [4]。 | Where enabled, admin-defined MCP hooks on the cloud coordinator for supported lifecycle and tool events; their coverage is to be checked separately from the Compliance API [5]. The governance page adds that analytics dashboards are not raw audit logs [6].在已启用的情况下，管理员定义的 MCP 钩子运行在云端协调器上，覆盖受支持的生命周期与工具事件；其覆盖范围需与合规 API 分开核对 [5]。治理页补充：分析仪表板不是原始审计日志 [6]。 | Agent instances are included in policies like users for data loss prevention, insider risk, communication compliance, eDiscovery and retention [9].智能体实例像用户一样被纳入数据丢失防护、内部风险、通信合规、电子取证与保留策略 [9]。 |
All cells are what the vendors’ own documents say; no feature was tested.
表中全部是厂商自己的文档所说，没有测试任何功能。
Key findings主要发现
| # | Finding · 发现发现 | Cite |
|---|---|---|
| F1 | Anthropic’s Activity Feed is retained 6 years and queryable within 1 minute, but recording starts only when the Compliance API is enabled and is not backfilled.Anthropic 的 Activity Feed 保留 6 年、1 分钟内可查询，但从启用合规 API 才开始记录，不回填。 | [1][3] |
| F2 | Anthropic session transcripts include tool-use and tool-result blocks, but tool inputs and results are cut to 10,000 bytes by default, and thinking blocks, images, the system prompt and tool definitions are absent.Anthropic 的会话记录含工具调用与工具结果块，但工具输入与结果默认被截断为 10,000 字节，且没有思考块、图片、系统提示词和工具定义。 | [2][3] |
| F3 | Anthropic’s Compliance API excludes prompt and response content from Console or API-key workloads, Claude Code run with a Console API key or through Bedrock, Google Cloud or Microsoft Foundry, and sessions under zero data retention.Anthropic 的合规 API 不含 Console 或 API 密钥工作负载的提示词与回应内容，不含用 Console API 密钥或经 Bedrock、Google Cloud、Microsoft Foundry 运行的 Claude Code，也不含零数据保留下的会话。 | [3][2] |
| F4 | OpenAI’s compliance logs are described as immutable and append-only, retained 30 days, for Enterprise and Edu workspaces; partners that ingest them include Microsoft Purview.OpenAI 的合规日志被描述为不可变、仅追加，保留 30 天，面向 Enterprise 与 Edu 工作区；接入它们的合作方包括 Microsoft Purview。 | [11] |
| F5 | OpenAI’s Codex compliance page says its sources do not establish a complete record of every local command, file operation, screenshot, approval or external action, and tells customers to run a representative task and compare the export with what was done.OpenAI 的 Codex 合规页面说其来源不能构成每条本地命令、文件操作、截图、批准或外部动作的完整记录，并建议客户跑一个有代表性的任务，把导出内容与实际做的动作对照。 | [5] |
| F6 | Microsoft’s audit record for AI interactions names the agent (ID, name, version), accessed resources with action and status, plugins and model provider, but its property table lists message IDs and flags rather than text.Microsoft 针对 AI 交互的审计记录写入智能体（ID、名称、版本）、被访问资源及动作与状态、插件与模型提供方，但属性表里列的是消息 ID 与标志而不是文本。 | [7] |
| F7 | Microsoft Purview audits an Agent 365 agent instance like a user, for agent-to-human, human-to-agent, agent-to-tools and agent-to-agent interactions.Microsoft Purview 像审计用户一样审计 Agent 365 的智能体实例，覆盖智能体与人、人与智能体、智能体与工具、智能体之间的交互。 | [9] |
| F8 | Purview Audit (Standard) keeps records 180 days; Audit (Premium) up to 1 year, and 10 years with an add-on licence; non-Microsoft AI app logs are pay-as-you-go and kept 180 days; ChatGPT Enterprise needs a connector scan and does not support sensitivity labels or data loss prevention.Purview Audit (Standard) 保留 180 天；Audit (Premium) 最长 1 年，加购附加许可可到 10 年；非 Microsoft AI 应用的日志按用量计费、保留 180 天；ChatGPT Enterprise 需要连接器扫描，且不支持敏感标签和数据丢失防护。 | [8][7][10] |
| F9 | Anthropic offers a push route next to the pull API: inference hooks can archive every governed transcript as it arrives while always returning allow (beta, Claude Enterprise, not on Bedrock or Google Cloud).Anthropic 在拉取 API 之外还有推送路径：推理钩子可以在每条受管对话到达时归档并始终返回允许（测试版，Claude Enterprise，Bedrock 与 Google Cloud 上不可用）。 | [4] |
Evidence notes证据说明
All sources are the vendors’ own documentation. The OpenAI Admin API reference, which the OpenAI pages name as the source of truth for event schemas and retention, was not read, so schema details and any retention differences by log type are unknown. The Anthropic session-transcript page is long and only its comparison table, coverage notes and retention section were read. The Microsoft pages were returned as full text by the fetch tool. Third-party descriptions of OpenAI log types (such as APP_LOG and CODEX_LOG) turned up in search results and were not used.
所有来源都是厂商自己的文档。OpenAI 的页面把 Admin API 参考称为事件模式与保留期的权威来源，这份参考没有读，所以各日志类型的字段细节和可能的保留期差异未知。Anthropic 会话记录页很长，只读了其中的对比表、覆盖说明和保留期部分。Microsoft 页面由抓取工具返回全文。搜索结果里出现了第三方对 OpenAI 日志类型的描述（如 APP_LOG、CODEX_LOG），没有采用。
Industry examples行业示例
| Company | Product | What it shows | Source |
|---|---|---|---|
| Anthropic | Compliance API | Long retention, near-real-time query and tool-call content, within the limits of F2 and F3.保留期长、近实时可查、含工具调用内容，但受 F2、F3 的限制。 | [1][2][3] |
| OpenAI | Compliance Logs Platform | Immutable export for SIEM, DLP and eDiscovery, with a 30-day window and a stated completeness caveat for local actions.面向 SIEM、DLP 与电子取证的不可变导出，保留窗口 30 天，并明说本地动作不保证完整。 | [11][5] |
| Microsoft | Purview Audit | A cross-vendor audit layer that treats agents like users, with shorter default retention and per-use billing for non-Microsoft apps.把智能体当作用户来审计的跨厂商审计层，默认保留期较短，非 Microsoft 应用按用量计费。 | [7][8][9][10] |
Architecture implications架构含义
The three vendors cover different places in the stack: Anthropic and OpenAI log what happens inside their own products; Microsoft Purview collects records across vendors and treats agents as identities. A customer with agents from several vendors would therefore pull vendor logs into its own archive and optionally into Purview, and would not rely on any single vendor’s retention window. This is my reading of the documents, not a design any vendor recommends.
三家厂商处在技术栈的不同位置：Anthropic 与 OpenAI 记录自己产品内部发生的事；Microsoft Purview 跨厂商收集记录，并把智能体当作身份。因此，使用多个厂商智能体的客户会把各厂商日志拉进自己的归档（也可选择送入 Purview），而不会只依赖某一家的保留窗口。这是我对文档的解读，不是任何厂商推荐的设计。
MDLC implicationsMDLC 含义
For an always-on agent with no release event (Q6), the action log is the main record. Of the three, only Anthropic’s documents describe tool calls and results as retained content with a 6-year window, and even there the default truncation and the exclusions in F3 mean a customer should measure what is missing. OpenAI states a completeness caveat itself. Microsoft’s agent audit shows that an interaction happened and which resources were touched, but the pages read do not show the content. None of the pages describes how an auditor would verify that a log was not altered.
对没有发版时刻的常驻智能体（Q6），动作日志是主要记录。三家里，只有 Anthropic 的文档把工具调用与结果描述为保留内容并给出 6 年窗口，即便如此，默认截断和 F3 的排除项也意味着客户应当实测缺了什么。OpenAI 自己写明了完整性方面的保留。Microsoft 的智能体审计能说明发生过交互、碰过哪些资源，但所读页面没有显示内容。没有任何页面描述审计师如何验证日志没有被改过。
My interpretation我的解读
No log here is “best” in general. For completeness of tool activity and retention, Anthropic’s documents are the strongest of the three, but only for Enterprise-signed-in surfaces. For a single audit layer across vendors, Microsoft Purview is the only candidate in this set, at the cost of shorter default retention and metadata-heavy records. OpenAI’s stated immutability is a useful property, but 30 days and an explicit completeness caveat mean it works as a feed into the customer’s own archive. This ranking would change if the OpenAI Admin API reference or tests show different coverage.
总体上这里没有哪一份日志是“最好”。就工具活动的完整性和保留期而言，Anthropic 的文档在三者中最强，但仅限 Enterprise 登录的界面。要跨厂商只用一个审计层，这组里只有 Microsoft Purview 可选，代价是默认保留期较短、记录偏元数据。OpenAI 声明的不可变性是有用的属性，但 30 天加上明说的完整性保留，使它更适合作为客户自有归档的数据源。如果 OpenAI Admin API 参考或实测显示的覆盖不同，这个排序会变。
Hypotheses假设
Regulated customers will keep their own archive of agent action records, with their own hashes, whichever vendor log they start from, because no vendor documents tamper evidence and retention windows range from 30 days to 6 years.
受监管客户不论从哪家厂商日志出发，都会自建带自有哈希的智能体动作记录归档，因为没有厂商记载防篡改机制，而且保留窗口从 30 天到 6 年不等。
Open questions未决问题
What do the OpenAI Admin API schemas and retention say per log type, and are dots’ actions in conversation_message or codex_log records? Do Anthropic transcripts for a long-running Managed Agents session appear in the Compliance API (the pages read cover Cowork and Claude Code, not Managed Agents)? Does the Purview audit record or activity explorer hold the prompt and response text, or only references? Do any of the three offer hashes, signatures or external timestamping? How complete is each export when compared with a known task (the test OpenAI suggests)?
OpenAI Admin API 的各日志类型字段与保留期是什么，dots 的动作在不在 conversation_message 或 codex_log 里？Anthropic 针对长时间运行的 Managed Agents 会话，其记录会不会出现在合规 API 里（所读页面覆盖的是 Cowork 与 Claude Code，不含 Managed Agents）？Purview 的审计记录或活动资源管理器里存的是提示词与回应文本，还是只有引用？三家有没有提供哈希、签名或外部时间戳？拿一个已知任务去比对，各自的导出有多完整（OpenAI 建议的那种测试）？
What would change my view?什么会改变我的看法？
- A vendor documenting hashes, signatures or external timestamps for audit exports.某厂商记载了审计导出的哈希、签名或外部时间戳。
- A hands-on comparison of exports against a known agent task showing different completeness from what the documents imply.对一个已知智能体任务的实测比对显示，完整性与文档所暗示的不同。
- OpenAI documenting retention longer than 30 days or dots coverage in its compliance logs.OpenAI 记载了长于 30 天的保留期，或把 dots 纳入其合规日志。
Link to hypotheses与假设的关系
Q2 (which systems of record own agent action logs): partly answered on what vendors document: each vendor owns its own log and Microsoft Purview is positioned as a cross-vendor layer; whether enterprises actually use it that way is not shown. Q6: informs what evidence exists, not how much is enough. H4: weak support (an audit layer across vendors exists in Microsoft’s documents). Not enough to change thesis confidence; it feeds the monthly review. The thesis page is not edited for a single note.
Q2（谁承载智能体动作日志）：就厂商文档所述已部分回答——各家拥有自己的日志，Microsoft Purview 被定位为跨厂商层；企业是否真的这样用，没有显示。Q6：说明了有哪些证据，没说明多少才够。H4：弱支持（Microsoft 的文档里存在跨厂商的审计层）。不足以改变论点置信度，会进入月度复查。论点页不为单篇笔记修改。
References参考文献
- Anthropic, “Query the Activity Feed,” Compliance API. https://platform.claude.com/docs/en/manage-claude/compliance-activity-feed vendor documentation · S-0132
- Anthropic, “Retrieve session transcripts,” Compliance API. Long page; the comparison table, coverage notes and retention section were read. https://platform.claude.com/docs/en/manage-claude/compliance-sessions vendor documentation · S-0133
- Anthropic, “Design your compliance integration.” https://platform.claude.com/docs/en/manage-claude/compliance-integration-patterns vendor documentation · S-0134
- Anthropic, “Inference hooks” (beta). https://platform.claude.com/docs/en/manage-claude/inference-hooks vendor documentation · S-0135
- OpenAI, “Compliance API and audit events” (Codex enterprise docs on ChatGPT Learn); read in full in a browser. https://learn.chatgpt.com/docs/enterprise/compliance-api vendor documentation · S-0136
- OpenAI, “Governance” (Codex enterprise docs on ChatGPT Learn); the first part was read. https://learn.chatgpt.com/docs/enterprise/governance vendor documentation (partial) · S-0137
- Microsoft Learn, “Audit logs for Copilot and AI applications” (ms.date 2026-08-26). https://learn.microsoft.com/en-us/purview/audit-copilot vendor documentation · S-0138
- Microsoft Learn, “Learn about auditing solutions in Microsoft Purview” (ms.date 2026-05-18). https://learn.microsoft.com/en-us/purview/audit-solutions-overview vendor documentation · S-0139
- Microsoft Learn, “Use Microsoft Purview to manage data security & compliance for Microsoft Agent 365.” https://learn.microsoft.com/en-us/purview/ai-agent-365 vendor documentation · S-0140
- Microsoft Learn, “Use Microsoft Purview to manage data security & compliance for ChatGPT Enterprise.” https://learn.microsoft.com/en-us/purview/ai-chatgpt-enterprise vendor documentation · S-0141
- OpenAI Help Center, “OpenAI Compliance Platform for Enterprise and Edu customers”; read in full (already registered). https://help.openai.com/en/articles/9261474-openai-compliance-platform-for-enterprise-and-edu-customers vendor documentation · S-0061
- Anthropic, “Compliance API” overview (already registered). https://platform.claude.com/docs/en/manage-claude/compliance-api vendor documentation · S-0111
- OpenAI dots documents (already registered); see the dots note. https://help.openai.com/en/articles/20001554-manage-dots-in-chatgpt-workspaces vendor documentation · S-0019 · S-0020
Last updated 2026-10-19.