How Amazon Bedrock catches AI-generated phishingAmazon Bedrock 如何捕获 AI 生成的网络钓鱼

Social engineering through phishing remains one of the most common tactics for launching cyberattacks. AI-generated phishing email messages now pose a new challenge for security teams managing email systems, significantly raising the risk because of their advanced sophistication. Modern social engineers use generative AI and open source intelligence (OSINT) to craft thousands of unique messages with perfect grammar, appropriate context, and personalized details. Today, an indicator of a phishing email message might be a perfectly written, professionally formatted message.通过网络钓鱼进行的社会工程攻击仍然是发起网络攻击最常用的手段之一。AI 生成的网络钓鱼电子邮件现在给管理邮件系统的安全团队带来了新的挑战，由于其高度复杂性，风险显著增加。现代社会工程攻击者利用生成式 AI 和开源情报（OSINT）来构造数千条语法完美、上下文贴切且包含个性化细节的独特消息。如今，网络钓鱼电子邮件的指标可能是一条书写完美、格式专业的消息。

The evolution of phishing网络钓鱼的演变

For someone like John, an IT security engineer at a mid-sized firm, the rules of phishing detection were once straightforward: flag the typos, catch the generic salutations, and quarantine anything with a mismatched sender domain. These were the defining characteristics of an earlier era of phishing, when attacks sent millions of generic, error-riddled email messages at scale, relying on volume rather than precision to find victims. Security filters were built exactly for these threats, and for years, they were effective. Poor grammar, generic greetings, and mismatched logos were indicators that gave attackers away.对于像 John 这样在一家中型公司工作的 IT 安全工程师来说，网络钓鱼检测的规则曾经很简单：标记拼写错误、识别通用称呼，并隔离任何发件人域名不匹配的邮件。这些是早期网络钓鱼时代的典型特征，当时攻击者大规模发送数百万封通用且错误百出的电子邮件，依靠数量而非精准度来寻找受害者。安全过滤器正是针对这些威胁而构建的，多年来一直有效。糟糕的语法、通用的问候语和不匹配的徽标都是暴露攻击者的指标。

The threat landscape John monitors today looks nothing like the ones those filters were designed to catch. Generative AI changed how phishing works. Attacks are now grammatically correct, contextually accurate, and personalized to the target. These messages don’t trigger traditional filters because those filters weren’t designed to catch them.John 今天监控的威胁环境与那些过滤器设计时所要捕获的完全不同。生成式 AI 改变了网络钓鱼的运作方式。如今的攻击语法正确、上下文准确，并且针对目标进行个性化定制。这些消息不会触发传统过滤器，因为那些过滤器并不是为捕获它们而设计的。

The threat is no longer identifiable by what it looks like, but what it knows. Modern AI systems run OSINT operations that pull data from professional networks, corporate websites, and publicly available digital footprints to map out organizational hierarchies and relationships. With that intelligence, social engineers can process massive datasets at scale to generate contextually accurate messages personalized to your organization. These communications can even adapt in real time based on your responses, shifting tone or adjusting details to stay consistent with the conversation.威胁不再通过其外在表现来识别，而是通过其掌握的信息来识别。现代 AI 系统执行 OSINT 操作，从专业网络、企业网站和公开可用的数字足迹中提取数据，以描绘组织架构和关系。有了这些情报，社会工程攻击者可以大规模处理海量数据集，生成针对您组织量身定制的上下文准确的消息。这些通信甚至可以根据您的回复实时调整，改变语气或调整细节，以保持对话的一致性。

Amazon Bedrock is a fully managed service that makes high-performing foundation models (FMs) from leading AI companies available through a unified API, along with capabilities needed to build generative AI applications with security, privacy, and responsible AI. Amazon Bedrock adds an additional layer of analysis to your existing security infrastructure that goes beyond traditional surface-level filtering. It understands context and detects phishing attempts based on behavioral patterns, not grammar quality or formatting. To put that into practice, let’s break down how Amazon Bedrock analyzes an email from the moment it hits your inbox.Amazon Bedrock 是一项完全托管的服务，通过统一 API 提供来自领先 AI 公司的高性能基础模型（FM），以及构建具有安全性、隐私性和负责任 AI 的生成式 AI 应用所需的功能。Amazon Bedrock 为您现有的安全基础设施增加了一层超越传统表面过滤的分析层。它能够理解上下文，并基于行为模式而非语法质量或格式来检测网络钓鱼尝试。为了付诸实践，让我们详细分析 Amazon Bedrock 如何从邮件进入收件箱的那一刻起对其进行分析。

Amazon Bedrock uses large-scale general-purpose AI models pre-trained on vast amounts of data. Foundation models can analyze behavioral patterns in email content, understand contextual relationships, and identify anomalies that signal a message might be a phishing attempt. In practice, these capabilities can be structured as a multi-stage analysis pipeline. Each email passes through authentication, behavior analysis, and risk scoring before reaching your users’ inboxes.Amazon Bedrock 使用在海量数据上预训练的大规模通用 AI 模型。基础模型可以分析电子邮件内容中的行为模式、理解上下文关系，并识别表明消息可能是网络钓鱼尝试的异常。在实践中，这些能力可以构建为多阶段分析管道。每封电子邮件在到达用户收件箱之前都要经过身份验证、行为分析和风险评估。

Amazon Bedrock offers two integrated capabilities to power your AI-driven phishing defense. Pre-trained foundation models bring sophisticated natural language understanding that can detect nuanced manipulation, contextual anomalies, and impersonation patterns invisible to rule-based systems. The second capability, Amazon Bedrock Guardrails, provides configurable safeguards that help align foundation model interactions with your organization’s responsible AI policies and application requirements, without requiring custom detection logic. Together, these capabilities can be integrated into a multi-stage email analysis pipeline.Amazon Bedrock 提供两项集成能力来支持您的 AI 驱动的网络钓鱼防御。预训练的基础模型带来复杂的自然语言理解能力，可以检测基于规则的系统无法察觉的微妙操纵、上下文异常和冒充模式。第二项能力是 Amazon Bedrock Guardrails，它提供可配置的安全防护，帮助将基础模型交互与您组织的负责任 AI 策略和应用需求保持一致，而无需自定义检测逻辑。这两种能力可以集成到一个多阶段电子邮件分析管道中。

Amazon Bedrock workflow for intelligent phishing defenseAmazon Bedrock 智能网络钓鱼防御工作流

In the workflow solution, each message first undergoes standard authentication checks (Sender Policy Framework (SPF), DomainKeys Identified Mail (DKIM), Domain-based Message Authentication, Reporting and Conformance (DMARC)). These protocols confirm that the sending server is authorized to send on behalf of the domain and that the message hasn’t been tampered with in transit. The phishing detection workflow, powered by the Amazon Bedrock foundation models, analyzes the message against three key factors: word choice, communication style deviations, and contextual appropriateness of requests. Detecting these subtle inconsistencies in writing style and misaligned requests adds a deeper layer of analysis on top of traditional security controls. AI analysis also requires careful governance to confirm it operates responsibly and within your defined boundaries. Amazon Bedrock Guardrails help filter both input prompts and model outputs. They prevent responses that could inadvertently leak confidential data, and they check that analysis results adhere to the policies you set. Keep in mind that guardrails need careful configuration and calibration to meet your application requirements.在工作流解决方案中，每封邮件首先接受标准身份验证检查（发件人策略框架 (SPF)、域名密钥识别邮件 (DKIM)、基于域的消息验证、报告与一致性 (DMARC)）。这些协议确认发送服务器已获授权代表该域名发送邮件，并且消息在传输过程中未被篡改。由 Amazon Bedrock 基础模型驱动的网络钓鱼检测工作流根据三个关键因素分析消息：用词选择、沟通风格偏差和请求的上下文适当性。检测这些写作风格上的细微不一致和错位的请求，可在传统安全控制之上增加更深入的分析层。AI 分析还需要谨慎治理，以确保其负责任地运行并在您定义的边界内操作。Amazon Bedrock Guardrails 有助于过滤输入提示和模型输出。它们防止可能无意中泄露机密数据的响应，并检查分析结果是否符合您设置的策略。请记住，防护措施需要仔细配置和校准，以满足您的应用需求。

Implementing Amazon Bedrock Guardrails for analysis实施 Amazon Bedrock Guardrails 进行分析

Amazon Bedrock Guardrails give you granular control over how foundation models process email content through content filters, denied topics, word filters, and sensitive information filters. For example, John the security engineer can configure guardrails to automatically redact sensitive personally identifiable information (PII) discovered during email analysis, helping to prevent the foundation model from generating responses that could inadvertently leak confidential data.Amazon Bedrock Guardrails 通过内容过滤器、拒绝的主题、单词过滤器和敏感信息过滤器，让您对基础模型如何处理电子邮件内容进行细粒度控制。例如，安全工程师 John 可以配置防护措施，自动编辑在电子邮件分析过程中发现的敏感个人身份信息（PII），帮助防止基础模型生成可能无意中泄露机密数据的响应。

However, guardrail configurations for security analysis require careful calibration. While content filters protect against inappropriate inputs and outputs, overly restrictive settings can prevent the model from analyzing suspicious content that legitimately needs to be evaluated. If a social engineer includes offensive language in an email message to bypass filters, your guardrails must allow the security system to analyze that content. At the same time, the guardrails must still protect against inappropriate inputs and outputs in other contexts. Guardrails also provide contextual grounding checks that keep model responses factually anchored to the email content being analyzed, reducing false positives caused by model hallucination. This allows the AI-powered analysis to operate within defined boundaries while still detecting intricate patterns.然而，用于安全分析的防护配置需要仔细校准。虽然内容过滤器可以防止不当的输入和输出，但过度严格的设置可能会阻止模型分析那些确实需要评估的可疑内容。如果社会工程攻击者在电子邮件中包含冒犯性语言以绕过过滤器，您的防护措施必须允许安全系统分析该内容。与此同时，防护措施仍必须防止其他上下文中的不当输入和输出。Guardrails 还提供上下文接地检查，使模型响应在事实上锚定于正在分析的电子邮件内容，从而减少由模型幻觉引起的误报。这使 AI 驱动的分析能够在定义范围内运行，同时仍能检测复杂的模式。

In this post, you will learn how to implement a multi-stage email analysis pipeline using Amazon Bedrock foundation models that evaluate sender behavior patterns, contextual appropriateness, and communication anomalies to identify AI-generated phishing attempts before they reach your users.在本文中，您将学习如何使用 Amazon Bedrock 基础模型实现多阶段电子邮件分析管道，该管道评估发件人行为模式、上下文适当性和通信异常，以便在 AI 生成的网络钓鱼尝试到达您的用户之前将其识别出来。

Implementation framework实施框架

The following framework shows how to put this into practice within your existing email security infrastructure, so that someone in John’s position can move from reactive filtering to proactive detection. After your standard authentication checks (SPF, DKIM, DMARC) confirm an email comes from a legitimate mail server, the phishing detection workflow goes a step further by layering in behavioral analysis. Your system moves from checking whether a server is authorized to evaluating whether a message matches how your coworker normally communicates.以下框架展示了如何在您现有的电子邮件安全基础设施中实践这一方法，使处于 John 这样职位的人能够从被动过滤转向主动检测。在您的标准身份验证检查（SPF、DKIM、DMARC）确认电子邮件来自合法邮件服务器后，网络钓鱼检测工作流通过引入行为分析更进一步。您的系统从检查服务器是否获得授权，转变为评估消息是否与您的同事通常的沟通方式相匹配。

Figure 1 maps the five-step email security analysis workflow, from initial guardrail screening through AI analysis, risk scoring, and final routing decisions.图 1 展示了五步电子邮件安全分析工作流，从初始防护筛选到 AI 分析、风险评估和最终路由决策。

Before diving into the implementation, let’s clarify what each component does. Behavioral analysis starts with a sender baseline tracker, which is a profile of each person who sends email to you. The sender baseline tracker logs how your employees normally write, so the Amazon Bedrock analysis pipeline has a reference point to compare against.在深入实施之前，让我们明确每个组件的作用。行为分析从发件人基线跟踪器开始，该跟踪器是每个向您发送电子邮件的人的画像。发件人基线跟踪器记录您的员工通常的写作方式，以便 Amazon Bedrock 分析管道有一个可供比较的参考点。

Over continued use, the phishing detection workflow will understand the words your employees use, how formal or casual they are, what they usually ask for, and who they normally communicate with. Consider John’s environment: A coworker who usually sends quick one-liners suddenly writes a formal email requesting an urgent wire transfer. The analysis pipeline catches that shift and flags it for John’s team to take a closer look.随着持续使用，网络钓鱼检测工作流将理解您的员工使用的词语、他们的正式或随意程度、他们通常要求什么，以及他们通常与谁沟通。以 John 的环境为例：一位通常发送简短一句话的同事突然写了一封正式电子邮件，要求紧急电汇。分析管道会发现这一变化，并标记出来供 John 的团队仔细检查。

This can help reduce false alarms and save time that John’s team might otherwise spend sorting through flagged email messages that turn out not to be real threats.这有助于减少误报，并节省 John 的团队本可能花在筛选最终并非真正威胁的标记邮件上的时间。

Here’s a high-level outline on how these components work together when an email enters your phishing detection workflow:以下是当电子邮件进入您的网络钓鱼检测工作流时，这些组件如何协同工作的高级概述：

Step 1: Input guardrails and pre-processing步骤 1：输入防护和预处理

INITIALIZE EmailSecurityAnalyzer:
    - Set up Amazon Bedrock client (Claude Sonnet 4.5 model)
    - Configure Amazon Bedrock Guardrails for PII protection and content filtering
    - Initialize knowledge base for phishing examples
    - Initialize sender baseline tracker
    - Set risk thresholds (safe < 30, suspicious < 70, dangerous >= 70)

FUNCTION analyze_email(email):
    // Step 1: Pre-process with guardrails
    processed_email = apply_input_guardrails(email)
    IF content_blocked:
        RETURN manual_review_required

The phishing detection workflow first runs incoming email messages through Amazon Bedrock Guardrails, which screen for sensitive content and flag anything that should go to manual review before the analysis begins.网络钓鱼检测工作流首先将传入的电子邮件消息通过 Amazon Bedrock Guardrails 运行，这些防护措施会筛选敏感内容，并在分析开始前标记任何应进入人工审核的内容。

Step 2: Prompt construction with context步骤 2：带上下文的提示构建

// Step 2: Build analysis prompt
prompt = construct_prompt(
    email_content,
    sender_baseline_patterns,
    organizational_context,
    known_phishing_examples
)

After an email clears that check, the workflow constructs an analysis prompt by combining the email’s content with the sender’s baseline communication patterns, organizational context, and known phishing examples by using Amazon Bedrock Knowledge Bases. That way, the model is evaluating the message against a full picture, not in a vacuum.在电子邮件通过该检查后，工作流通过结合电子邮件的内容与发件人的基线沟通模式、组织上下文以及已知的网络钓鱼示例，并利用 Amazon Bedrock Knowledge Bases 构建分析提示。这样，模型就能在完整背景下评估消息，而不是在真空中进行。

Step 3: AI-powered analysis with guardrails步骤 3：带防护的 AI 驱动分析

// Step 3: Invoke AI model with guardrails
analysis = bedrock_invoke_with_guardrails(prompt)
IF guardrail_intervened:
    RETURN blocked_with_reasons

The foundation model processes the email using the constructed prompt while guardrails keep the analysis within your defined security boundaries. The foundation model can examine suspicious content thoroughly while the guardrails keep it from generating outputs that expose sensitive information in the process.基础模型使用构建的提示处理电子邮件，同时防护措施使分析保持在您定义的安全边界内。基础模型可以彻底检查可疑内容，而防护措施则防止其生成在此过程中暴露敏感信息的输出。

Step 4: Multi-factor risk scoring步骤 4：多因素风险评分

// Step 4: Calculate risk scores
risk_score = weighted_average(
    content_anomaly_score,
    behavioral_deviation_score,
    context_alignment_score
)

From that analysis, the Amazon Bedrock pipeline generates three scores: one for content anomalies, one for behavioral deviations, and one for contextual alignment. The pipeline combines them into a single risk score from 0–100, which determines where the email is routed.根据该分析，Amazon Bedrock 管道生成三个分数：一个用于内容异常，一个用于行为偏差，一个用于上下文一致性。管道将它们合并为一个 0–100 的风险评分，该评分决定了电子邮件的路由位置。

Step 5: Classification and automated routing步骤 5：分类和自动路由

// Step 5: Classify and route
risk_level = classify_risk(risk_score)
action = route_email(risk_level) // DELIVER, QUARANTINE, or BLOCK
RETURN analysis_result

FUNCTION route_email(risk_level):
    IF risk_level == SAFE: deliver_to_inbox
    IF risk_level == SUSPICIOUS: quarantine_for_review
    IF risk_level == DANGEROUS: block_and_alert_security

Safe messages land in your employees’ inboxes as usual. Suspicious email messages get quarantined for your security team to review. Dangerous messages are blocked outright.安全消息照常进入您的员工的收件箱。可疑的电子邮件消息被隔离，供您的安全团队审查。危险消息被直接阻止。

Continuous learning through feedback通过反馈持续学习

FUNCTION process_feedback(email, is_phishing):
    IF is_phishing:
        add_to_phishing_knowledge_base(email)
    ELSE:
        update_sender_baseline(email)
        add_to_legitimate_examples(email)

These steps happen in milliseconds as messages move through your routing system. Your existing infrastructure still handles message routing and delivery. The analysis runs alongside it as an inspection layer that evaluates behavioral risk before messages reach your users’ inboxes.这些步骤在消息通过您的路由系统时以毫秒级完成。您现有的基础设施仍处理消息路由和传递。分析作为一个检查层与之并行运行，在消息到达您的用户收件箱之前评估行为风险。

Over continued use, the phishing detection workflow improves its accuracy in making these calls through a few complementary techniques. Dynamic prompt engineering, the practice of iteratively refining the instructions sent to the foundation model based on real-world results, takes feedback from the security team and incorporates it directly into your analysis prompts, gradually fine-tuning how the model evaluates potential issues. That feedback loop also feeds into a growing knowledge base of validated examples, where confirmed phishing attempts and legitimate messages are cataloged and later used as few-shot learning demonstrations in future prompts. So, when a new email comes in, the model isn’t working from scratch. It references your real, previously verified examples that match similar patterns to make a more informed judgment.随着持续使用，网络钓鱼检测工作流通过几种互补技术提高了做出这些判断的准确性。动态提示工程是一种基于实际结果迭代优化发送给基础模型的指令的做法，它接收安全团队的反馈并将其直接纳入您的分析提示中，逐步微调模型评估潜在问题的方式。该反馈循环还输入到一个不断增长的已验证示例知识库中，其中确认的网络钓鱼尝试和合法消息被编目，并在以后的提示中用作少样本学习演示。因此，当新电子邮件到达时，模型不是从零开始。它会参考与类似模式匹配的真实的、先前验证过的示例，以做出更明智的判断。

Example: AI-generated phishing email analysis示例：AI 生成的网络钓鱼电子邮件分析

The following AI-generated phishing email message demonstrates modern phishing sophistication. Notice the perfect grammar, legitimate business context, and reference to a real purchase order (PO) format. None of these would trigger traditional spam filters. Following the email message is a simplified prompt structure showing how Amazon Bedrock analyzes messages against sender baselines and known phishing patterns. The prompt combines email content with historical context to support behavioral analysis beyond surface-level filtering. Last is a sample risk assessment output identifying a vendor impersonation attempt. The Amazon Bedrock pipeline flagged behavioral anomalies, including a first-ever payment change request, along with domain inconsistencies that traditional authentication checks missed.以下 AI 生成的网络钓鱼电子邮件消息展示了现代网络钓鱼的复杂性。请注意其完美的语法、合法的业务上下文以及对真实采购订单（PO）格式的引用。这些都不会触发传统的垃圾邮件过滤器。紧随电子邮件消息之后的是一个简化的提示结构，展示 Amazon Bedrock 如何根据发件人基线和已知网络钓鱼模式分析消息。该提示将电子邮件内容与历史上下文相结合，以支持超越表面过滤的行为分析。最后是一个示例风险评估输出，识别出供应商冒充尝试。Amazon Bedrock 管道标记了行为异常，包括首次出现的付款变更请求，以及传统身份验证检查遗漏的域名不一致。

Sample phishing email示例网络钓鱼电子邮件

Hi Sarah,

Following up on our last call Tuesday about the Q3 reconciliation.

Our finance team has updated our banking details as part of our transition to Example Banking Inc.

Could you update the payment info for PO-2024-089? Before the November 15th deadline? New details attached.

Best,
Michael Chen | Example Inc.

Prompt structure and risk assessment output提示结构和风险评估输出

=== EMAIL CONTENT ===
{email_content}

=== SENDER BASELINE ===
- Domain: example.com (verified vendor)
- History: 2-3 emails/month, never requested payment changes
- Tone: Professional, invoice/contract discussions

=== KNOWN EVENT PATTERNS ===
- Vendor impersonation with lookalike domains
- Payment detail change requests referencing valid POs risk assessment

=== Task ===
Score (0-100): content anomalies, behavioral deviation, context alignment

{
    "risk_score": 78,
    "risk_level": "DANGEROUS",
    "key_findings": [
        "Domain mismatch: 'example-website.com' vs 'example.com'",
        "First-ever payment change request from this sender",
        "Phone number doesn't match vendor records"
    ]
}

The continuous feedback loop持续反馈循环

Behind these examples, the phishing detection system maintains dynamic sender baselines in a database that tracks each of your sender’s typical communication patterns, vocabulary, tone, and request types. False positives flagged by John’s security team are fed back into the phishing detection pipeline, updating baselines to account for legitimate variations in how senders communicate. Confirmed phishing patterns are cataloged alongside these baselines to enrich future prompt context with current intelligence. The result is a feedback loop where every correction and every confirmed threat make the analysis more accurate.在这些示例的背后，网络钓鱼检测系统在数据库中维护动态发件人基线，跟踪每个发件人的典型沟通模式、词汇、语气和请求类型。John 的安全团队标记的误报会被反馈到网络钓鱼检测管道中，更新基线以考虑发件人沟通方式的合法变化。已确认的网络钓鱼模式与这些基线一起编目，以用最新情报丰富未来的提示上下文。结果形成了一个反馈循环，每一次修正和每一个已确认的威胁都使分析更加准确。

The continuous feedback pipeline runs across five stages:持续反馈管道分五个阶段运行：

1. Analyze – The foundation model evaluates your incoming email messages using dynamic prompts built from accumulated phishing attempt intelligence and sender context.1. 分析 – 基础模型使用根据累积的网络钓鱼尝试情报和发件人上下文构建的动态提示来评估您传入的电子邮件消息。

2. Score – Based on that analysis, a risk score from 0–100 is assigned, and suspicious messages are quarantined for your security team’s review.2. 评分 – 根据该分析，分配一个 0–100 的风险评分，可疑消息被隔离以供您的安全团队审查。

3. Review – Flagged messages get classified as either a confirmed phishing attempt or a false positive.3. 审查 – 被标记的消息被分类为已确认的网络钓鱼尝试或误报。

4. Learn – Those classifications feed back into your system, updating the example library, sender behavior baselines, and emerging patterns catalog.4. 学习 – 这些分类反馈到您的系统中，更新示例库、发件人行为基线和新兴模式目录。

5. Enhance – New examples and confirmed phishing attempt patterns get incorporated into the analysis prompts, improving detection accuracy for the next cycle.5. 增强 – 新的示例和已确认的网络钓鱼尝试模式被纳入分析提示中，提高下一个周期的检测准确性。

Early cycles will require more hands-on review as your system creates its baseline understanding. For John, that means his team initially spends more time classifying flagged messages, but the investment pays off quickly. As the example library and sender profiles grow, the model becomes progressively more accurate at distinguishing legitimate communications from phishing attempts. John stays in the loop throughout, but his attention shifts from sifting through noise to focusing on genuinely suspicious messages.在早期循环中，由于您的系统正在建立基线理解，将需要更多动手审查。对 John 来说，这意味着他的团队最初会花费更多时间对标记的消息进行分类，但这种投入很快会得到回报。随着示例库和发件人画像的增长，模型在区分合法通信与网络钓鱼尝试方面变得越来越准确。John 始终参与其中，但他的注意力从筛选噪音转移到关注真正可疑的消息上。

Each cycle through this loop creates a stronger, more adaptive defense that evolves alongside the phishing attempts it was designed to catch. That continuous improvement is what separates this feedback-driven detection model from static, signature-based detection.该循环的每一轮都会创造更强、更具适应性的防御，并随着它旨在捕获的网络钓鱼尝试而不断发展。这种持续改进正是这种反馈驱动的检测模型与静态的基于签名的检测之间的区别所在。

Conclusion结论

Phishing detection can no longer rely on surface-level indicators such as typos and awkward phrasing. The framework in this post addresses that shift by combining the Amazon Bedrock foundation models with behavioral analysis, contextual grounding, and a continuous feedback loop that improves accuracy over time. Amazon Bedrock catches subtle manipulation attempts that trained eyes might miss, while your existing infrastructure keeps doing what it was built to do.网络钓鱼检测不能再依赖于拼写错误和笨拙措辞等表面指标。本文中的框架通过将 Amazon Bedrock 基础模型与行为分析、上下文接地以及随时间提高准确性的持续反馈循环相结合，应对了这一转变。Amazon Bedrock 能够捕获训练有素的眼睛可能错过的微妙操纵尝试，而您现有的基础设施则继续执行其设计的功能。

Pair these defenses with solid verification processes, healthy skepticism toward unexpected requests, and a security culture that keeps your teams moving confidently. Employee awareness still matters, but now generative AI works with you to identify and help prevent impersonation attempts. AI made phishing harder to detect. The same technology, applied defensively, makes it harder to succeed.将这些防御措施与可靠的验证流程、对意外请求的合理怀疑以及让您的团队保持自信前行的安全文化相结合。员工意识仍然重要，但现在生成式 AI 与您合作，识别并帮助防止冒充企图。AI 让网络钓鱼更难被检测。同样的技术，如果用于防御，会让它更难得逞。

To begin implementing these defenses, start by visiting the Amazon Bedrock console. You can configure Amazon Bedrock Guardrails for your email flow and follow this tutorial to build your own email phishing detection pipeline. Share your experience with AI-powered security in the comments.要开始实施这些防御措施，请先访问 Amazon Bedrock 控制台。您可以为您的电子邮件流配置 Amazon Bedrock Guardrails，并按照本教程构建自己的电子邮件网络钓鱼检测管道。欢迎在评论中分享您对 AI 驱动安全性的体验。

About the authors关于作者