AWS AIP-C01 开源练习 · 高难 · 第 03 套

A. Invocation logging records full prompts, responses, and metadata to S3 and/or CloudWatch with no custom instrumentation.

B. CloudTrail records API metadata, not the full prompt/response content.

C. Metrics are numeric aggregates, not request/response content.

D. X-Ray traces request flow and timing, not full model content.

A. 调用日志记录将完整的提示、响应和元数据记录到 S3 和/或 CloudWatch，无需自定义工具。

B. CloudTrail 记录 API 元数据，而不是完整的提示/响应内容。

C. 指标是数字聚合，而不是请求/响应内容。

D. X-Ray 跟踪请求流和时间，而不是完整的模型内容。

A. CloudTrail records API calls with principal, action, time, and source, the audit trail of who did what.

B. Invocation logging captures model prompt/response content, not a full account-wide API caller audit trail.

C. Dashboards visualize metrics, not per-call caller identity.

D. Config tracks resource configuration compliance, not API caller activity.

A. CloudTrail 记录 API 调用的主体、操作、时间和来源，以及谁做了什么的审计跟踪。

B. 调用日志记录捕获模型提示/响应内容，而不是完整的帐户范围内的 API 调用者审核跟踪。

C. 仪表板可视化指标，而不是每次呼叫的呼叫者身份。

D. Config 跟踪资源配置合规性，而不是 API 调用者活动。

A. Knowledge base logging records per-file ingestion status, which Logs Insights can query for failures.

B. Invocation logging captures model calls, not the document ingestion process.

C. CloudTrail records API calls, not document-level ingestion status.

D. Application Signals monitors application performance, not knowledge base ingestion.

A. 知识库日志记录记录每个文件的摄取状态，Logs Insights 可以查询失败情况。

B. 调用日志记录捕获模型调用，而不是文档摄取过程。

C. CloudTrail 记录 API 调用，而不是文档级摄取状态。

D. 应用程序信号监视应用程序性能，而不是知识库摄取。

A. X-Ray traces requests across components and shows segment timings that pinpoint the bottleneck.

B. Alarms detect that latency is high but do not localize which component is responsible.

C. Invocation logging captures model content, not cross-service timing.

D. Macie classifies data and has no tracing capability.

A. X-Ray 跟踪组件之间的请求并显示可查明瓶颈的分段时序。

B. 警报检测到延迟很高，但无法定位是哪个组件负责。

C. 调用日志记录捕获模型内容，而不是跨服务计时。

D. Macie 对数据进行分类，并且没有跟踪功能。

A. Logs Insights provides an interactive query language to search and aggregate the CloudWatch log data.

B. CloudTrail Lake queries CloudTrail API events, not arbitrary invocation logs already in CloudWatch Logs.

C. Athena would require exporting and cataloging the data first; Logs Insights queries the logs directly.

D. QuickSight Q answers over prepared BI datasets, not ad hoc CloudWatch log querying.

A. Logs Insights 提供交互式查询语言来搜索和聚合 CloudWatch 日志数据。

B. CloudTrail Lake 查询 CloudTrail API 事件，而不是 CloudWatch Logs 中已有的任意调用日志。

C. Athena 需要首先导出数据并对其进行编目； Logs Insights 直接查询日志。

D. QuickSight Q 对准备好的 BI 数据集进行回答，而不是临时 CloudWatch 日志查询。

A. Model Monitor continuously checks a deployed endpoint for data quality and drift against a baseline and can alert.

B. A one-time Clarify analysis measures bias/explainability but is not continuous endpoint drift monitoring.

C. Infrastructure metrics do not detect data drift.

D. Guardrails filter content; they do not monitor endpoint data drift.

A. 模型监视器持续检查已部署端点的数据质量和相对于基线的偏差，并可以发出警报。

B. 一次性澄清分析衡量偏差/可解释性，但不是连续的终点漂移监测。

C. 基础设施指标不会检测数据漂移。

D. 护栏过滤内容；他们不监控端点数据漂移。

A. Application inference profiles tag and route invocations so usage and cost can be attributed per application or team.

B. Cost Explorer alone cannot split a shared model's spend across applications without attribution tags.

C. CloudTrail records who called the API, not per-application cost attribution.

D. Invocation logging stores content, not cost attribution.

A. 应用程序推理配置文件标记和路由调用，因此可以将使用情况和成本归因于每个应用程序或团队。

B. 如果没有归因标签，仅 Cost Explorer 无法在应用程序之间分配共享模型的支出。

C. CloudTrail 记录谁调用了 API，而不是每个应用程序的成本归因。

D. 调用日志记录存储内容，而不是成本归因。

A. Response logging plus scheduled automated evaluation on samples gives continuous, data-driven quality monitoring.

B. Infrastructure metrics do not reveal generation-quality regressions.

C. Relying on complaints is reactive and lets defects reach users.

D. A one-time test misses regressions introduced by later changes.

A. 响应记录加上对样品的预定自动评估可提供持续的、数据驱动的质量监控。

B. 基础设施指标并未揭示发电质量的回归。

C. 依赖投诉是被动的，会让缺陷传达给用户。

D. 一次性测试会错过后续更改引入的回归。

A. Identity pools federate authenticated users to temporary AWS credentials tied to an IAM role, with no embedded keys.

B. A user pool authenticates users but does not itself vend AWS credentials; the identity pool provides AWS credentials.

C. Embedding long-lived keys in a distributed app is a serious security risk.

D. IAM users issue long-lived keys and do not scale or provide temporary access.

A. 身份池将经过身份验证的用户联合到与 IAM 角色绑定的临时 AWS 凭证，没有嵌入密钥。

B. 用户池对用户进行身份验证，但本身并不出售 AWS 凭证；身份池提供 AWS 凭证。

C. 在分布式应用程序中嵌入长期密钥是一个严重的安全风险。

D. IAM 用户颁发长期密钥，不扩展或提供临时访问。

A. IAM Identity Center is the workforce SSO service that federates corporate identities to temporary AWS access.

B. Cognito targets application/customer identities; workforce SSO via the corporate IdP is IAM Identity Center.

C. IAM users rely on long-lived keys and do not integrate with the corporate IdP.

D. Distributing long-lived keys is insecure and unmanageable.

A. IAM Identity Center 是一种员工 SSO 服务，可将企业身份与临时 AWS 访问联合起来。

B. Cognito 针对应用程序/客户身份；通过企业 IdP 进行的员工 SSO 是 IAM 身份中心。

C. IAM 用户依赖长期密钥，并且不与企业 IdP 集成。

D. 分发长期密钥是不安全且难以管理的。

A. Interface endpoints via PrivateLink keep traffic to Bedrock on the AWS network.

B. A NAT gateway still sends traffic over the public internet.

C. TLS encrypts traffic but it still traverses the public internet.

D. VPC peering connects VPCs to each other, not your VPC privately to the Bedrock service.

A. 通过 PrivateLink 的接口终端节点将流量保留到 AWS 网络上的 Bedrock。

B. NAT 网关仍然通过公共互联网发送流量。

C. TLS 对流量进行加密，但它仍然穿过公共互联网。

D. VPC 对等互连将 VPC 相互连接，而不是将您的 VPC 私下连接到 Bedrock 服务。

A. Cross-account role assumption provides scoped, temporary credentials without sharing long-lived keys.

B. Sharing long-lived keys is insecure and unmanageable.

C. Public access removes the required access control.

D. Sharing root credentials is a severe security violation.

A. 跨账户角色假设提供了有范围的临时凭证，而无需共享长期密钥。

B. 共享长期密钥是不安全且难以管理的。

C. 公共访问消除了所需的访问控制。

D. 共享 root 凭据是严重的安全违规行为。

A. Secrets Manager stores credentials with IAM-controlled access and supports built-in automatic rotation.

B. Parameter Store can hold encrypted values but does not provide built-in managed rotation for arbitrary third-party secrets the way Secrets Manager does.

C. Plaintext secrets are insecure and cannot be rotated safely.

D. Hard-coded secrets cannot be rotated and risk exposure.

A. Secrets Manager 通过 IAM 控制的访问来存储凭证，并支持内置自动轮换。

B. Parameter Store 可以保存加密值，但不像 Secrets Manager 那样为任意第三方机密提供内置托管轮换。

C. 明文秘密是不安全的并且无法安全地轮换。

D. 硬编码的秘密无法轮换并存在暴露的风险。

A. Scoping the policy to the specific model ARN technically restricts the role to that model.

B. That condition enforces that a guardrail is applied; it does not restrict which model the role may invoke.

C. A verbal rule is not a technical control.

D. Application-side selection is not an access control and can be bypassed.

A. 将策略范围限定为特定模型 ARN 从技术上讲将角色限制为该模型。

B. 该条件强制应用护栏；它不限制角色可以调用哪个模型。

C. 口头规则不是技术控制。

D. 应用程序端选择不是访问控制，可以绕过。

A. A customer managed key gives control over key policy and rotation, and CloudTrail audits its use.

B. AWS owned keys encrypt data but give the customer no control, rotation policy, or usage auditing.

C. Unencrypted data fails the requirement.

D. A hard-coded passphrase is insecure and not an auditable managed key.

A. 客户管理的密钥可以控制密钥策略和轮换，CloudTrail 会审核其使用情况。

B. AWS 拥有的密钥可加密数据，但不给予客户任何控制权、轮换策略或使用情况审核。

C. 未加密的数据不符合要求。

D. 硬编码的密码是不安全的，并且不是可审核的托管密钥。

A. Bedrock keeps inference data in the Region you call and does not use customer inputs/outputs to train base models.

B. No opt-out is required; Bedrock does not use your data to train base models.

C. Inference data is processed in the Region you invoke, not globally.

D. Using Bedrock already prevents your data from training base models; self-hosting is unnecessary for that guarantee.

A. Bedrock 将推理数据保留在您调用的区域中，并且不使用客户输入/输出来训练基础模型。

B. 无需选择退出； Bedrock 不会使用您的数据来训练基础模型。

C. 推理数据在您调用的区域中处理，而不是全局处理。

D. 使用 Bedrock 已经可以防止您的数据训练基础模型；为了保证这一点，自托管是不必要的。

A. Content filters detect and block harmful categories in input and output with per-category strength settings.

B. Denied topics block specific subjects you define, not the built-in harmful-content categories.

C. Word filters block specific words/phrases, not graded harmful-content categories.

D. Stop sequences end generation; they do not classify harmful content.

A. 内容过滤器通过每个类别的强度设置来检测和阻止输入和输出中的有害类别。

B. 拒绝的主题会阻止您定义的特定主题，而不是内置的有害内容类别。

C. 单词过滤器会阻止特定单词/短语，而不是对有害内容类别进行分级。

D. 停止序列结束生成；他们不对有害内容进行分类。

A. Denied topics block a subject by semantic description and examples, catching varied phrasings.

B. A keyword list misses rephrasings and is brittle for a whole subject.

C. Content filters target harmful categories, not an arbitrary business subject.

D. The PII filter handles personal data, not topic restriction.

A. 被拒绝的主题通过语义描述和示例来阻止主题，捕捉不同的短语。

B. 关键字列表缺少重新措辞，并且对于整个主题来说是脆弱的。

C. 内容过滤器针对有害类别，而不是任意业务主题。

D. PII 过滤器处理个人数据，而不是主题限制。

A. The PII filter detects and masks or blocks personal information in inputs and outputs.

B. Comprehend can detect PII but is a separate custom step, not the managed in-guardrail control the scenario asks for.

C. Denied topics block subjects, not PII entities.

D. Content filters handle harmful categories, not PII masking.

A. PII 过滤器检测并屏蔽或阻止输入和输出中的个人信息。

B. Comprehend 可以检测 PII，但它是一个单独的自定义步骤，而不是场景要求的托管护栏内控制。

C. 被拒绝的主题会阻止主题，而不是 PII 实体。

D. 内容过滤器处理有害类别，而不是 PII 屏蔽。

A. The prompt attack filter flags prompt injection and jailbreak inputs so they can be blocked.

B. Denied topics block subjects, not injection attempts.

C. Those categories do not target instruction-override attacks.

D. Temperature does not detect prompt attacks.

A. 提示攻击过滤器标记提示注入和越狱输入，以便可以阻止它们。

B. 被拒绝的主题会阻止受试者，而不是注射尝试。

C. 这些类别不针对指令覆盖攻击。

D. 温度不检测即时攻击。

A. Word filters block specific, exact terms such as competitor names and profanity.

B. Denied topics target whole semantic subjects, which is heavier than blocking an exact word list.

C. Content filters address harmful categories, not a specific custom term list.

D. The PII filter handles personal data, not competitor names.

A. 单词过滤器会阻止特定、准确的术语，例如竞争对手的名称和脏话。

B. 拒绝的主题针对整个语义主题，这比阻止确切的单词列表更重。

C. 内容过滤器针对的是有害类别，而不是特定的自定义术语列表。

D. PII 过滤器处理个人数据，而不是竞争对手名称。

A. These checks score grounding in the source and relevance to the query and can block low-scoring responses at a configured threshold.

B. Denied topics block subjects, not ungrounded answers within an allowed subject.

C. Content filters target harmful content, not grounding.

D. Word filters block specific terms, not ungrounded responses.

A. 这些检查对源的基础和与查询的相关性进行评分，并可以在配置的阈值下阻止低分响应。

B. 被拒绝的主题会阻止主题，而不是允许主题内的无根据的答案。

C. 内容过滤器针对的是有害内容，而不是接地内容。

D. 单词过滤器会阻止特定术语，而不是没有根据的响应。

A. ApplyGuardrail evaluates input/output text against a guardrail independently of a Bedrock model call, covering external models and arbitrary points.

B. Inline-only guardrails cannot cover text from a non-Bedrock model or arbitrary pipeline steps.

C. Denied topics is a policy type, not the independent evaluation API.

D. Provisioned Throughput is a capacity feature, unrelated to guardrail evaluation.

A. ApplyGuardrail 根据护栏评估输入/输出文本，独立于 Bedrock 模型调用，涵盖外部模型和任意点。

B. 仅内联护栏无法覆盖非基岩模型或任意管道步骤中的文本。

C. 拒绝主题是一种策略类型，而不是独立评估 API。

D. 预置吞吐量是一项容量功能，与护栏评估无关。

A. Automated Reasoning checks use formal, logic-based verification against an encoded policy to validate responses.

B. Grounding checks whether a response is supported by provided source text, not formal verification against encoded policy rules.

C. Content filters handle harmful content, not policy conformance verification.

D. Temperature does not verify outputs against policy.

A. 自动推理检查使用针对编码策略的基于逻辑的正式验证来验证响应。

B. 接地检查所提供的源文本是否支持响应，而不是根据编码的策略规则进行正式验证。

C. 内容过滤器处理有害内容，而不是策略一致性验证。

D. 温度不会根据策略验证输出。

A. The condition key centrally and efficiently enforces that the guardrail is applied across all relevant API calls.

B. A proxy is a bottleneck and custom code to maintain, less efficient than the IAM condition.

C. Per-call retrieval adds latency and maintenance.

D. Per-app configuration is not centrally enforced and is easy to miss.

A. 条件键集中有效地强制在所有相关 API 调用中应用护栏。

B. 代理是一个瓶颈，需要维护自定义代码，效率低于 IAM 条件。

C. 每次调用检索会增加延迟和维护。

D. 每个应用程序的配置不是集中执行的，很容易被遗漏。

A. Compliance mode enforces WORM immutability that no one, including root, can override until retention expires.

B. Governance mode can be overridden by users with special permissions, so it does not guarantee immutability against administrators.

C. Versioning preserves versions but does not prevent deletion or enforce retention.

D. A bucket policy can be changed by administrators, so it is not immutable retention.

A. 合规性模式强制执行 WORM 不变性，在保留到期之前，任何人（包括根用户）都无法覆盖该不变性。

B. 治理模式可以被具有特殊权限的用户覆盖，因此它不能保证管理员的不变性。

C. 版本控制会保留版本，但不会阻止删除或强制保留。

D. 管理员可以更改存储桶策略，因此它不是不可变的保留。

A. CloudTrail records API calls with principal, action, and time.

B. Invocation logging captures prompt/response content, not the account-wide API caller audit trail.

C. Metrics are numeric aggregates, not caller identity.

D. Config tracks resource configuration, not API caller activity.

A. CloudTrail 记录 API 调用的主体、操作和时间。

B. 调用日志记录捕获提示/响应内容，而不是帐户范围的 API 调用者审核跟踪。

C. 指标是数字聚合，而不是呼叫者身份。

D. Config 跟踪资源配置，而不是 API 调用者活动。

A. Invocation logging captures full content, and Object Lock makes the stored logs immutable for retention.

B. CloudTrail captures API metadata, not the full prompt/response content.

C. Console output is not durable, complete, or immutable.

D. Metrics do not store prompt or response text.

A. 调用日志记录捕获完整内容，对象锁定使存储的日志不可变以便保留。

B. CloudTrail 捕获 API 元数据，而不是完整的提示/响应内容。

C. 控制台输出不持久、不完整或不可变。

D. 指标不存储提示或响应文本。

A. Detecting and redacting PII before storage keeps useful analytics data while complying with the no-PII rule.

B. Storing raw PII violates the requirement.

C. Encryption protects data at rest but does not satisfy a rule against retaining PII at all.

D. Temperature does not remove PII from stored text.

A. 在存储之前检测和编辑 PII 可保留有用的分析数据，同时遵守无 PII 规则。

B. 存储原始 PII 违反了要求。

C. 加密可以保护静态数据，但根本不满足保留 PII 的规则。

D. 温度不会从存储的文本中删除 PII。

A. Customer managed keys give control and rotation, and CloudTrail records key usage for auditing.

B. Unencrypted sensitive data fails the requirement.

C. A hard-coded shared secret is insecure and not auditable.

D. Public storage exposes the data and is the opposite of the requirement.

A. 客户管理的密钥提供控制和轮换，CloudTrail 记录密钥使用情况以供审核。

B. 未加密的敏感数据不符合要求。

C. 硬编码的共享秘密是不安全且不可审计的。

D. 公共存储暴露了数据，与需求相反。

A. Embeddings persist in the vector store until the knowledge base is updated, so both the source and the ingested vectors must be removed.

B. Deleting the source leaves the already-ingested embeddings in the vector store, so the data is not fully removed.

C. Logs are unrelated to the embeddings retained in the vector store.

D. Object Lock prevents deletion during retention, which is the opposite of fulfilling an erasure request.

A. 嵌入会一直保留在向量存储中，直到知识库更新为止，因此必须删除源向量和摄取的向量。

B. 删除源会将已提取的嵌入留在向量存储中，因此数据并未完全删除。

C. 日志与向量存储中保留的嵌入无关。

D. 对象锁定可防止在保留期间删除，这与满足擦除请求相反。