11-19 更新 实时
首页/ AI 大模型/ 正文

不是是否,而是如何:AI 攻击者时代的安全事件响应

AI 大模型 来源:MarkTechPost 2026-11-19 18:45 0 阅读 机器翻译

攻击者如今手中握有强大的 AI。AI 智能体能够以机器速度进行扫描、规划和实施入侵,而且每个月都在进步。结果是攻击更多、更频繁,并且呈全球规模。在这样的环境中,会有更多公司在某个时刻遭遇安全事件。真正拉开差距的,是它们如何应对。面对机器速度的威胁,需要一套包含 4 大支柱的现代化应对手册。自动检测与遏制。快速且安全地沟通。彻底清除。持续缩小影响范围。本文通过 6 个月的事件展示这一趋势,随后分析几家公司的案例,其中 Fireworks AI 是成熟且结构稳健的应对典范。

AI 改变了攻击的经济账。过去需要熟练人员耗费数天完成的任务,如今可以交给智能体执行。

11 月,Google Threat Intelligence Group 和 Mandiant 表示,AI 滥用已达到“关键拐点”。他们报告称,攻击者利用 AI 缩短时间线,并以机器速度实施多阶段入侵。CrowdStrike 报告称,有 1 个团伙在 11 天内利用 AI 智能体攻击了 200 多个组织。Palo Alto Networks 的 Unit 42 发现,攻击速度比一年前快了约 4 倍。

这一转变更早就已开始。2025 年,Anthropic 阻断了一场攻击行动,其中 80% 到 90% 的战术工作由 AI 完成,人类介入极少。

工具不断升级,因此节奏只会继续加快。AI 基础设施是格外有吸引力的目标,因为 AI 访问权限本身就有价值。过去 6 个月的事件展现了这种局面。

事件出现在 AI 技术栈的每一层。这些事件共同表明,相关攻击活动的覆盖面已变得多么广泛。

5月:软件包注册表。5月10日和11日,“Mini Shai-Hulud”攻击活动篡改了约170个npm软件包以及部分PyPI库。受影响的项目包括Mistral AI和Guardrails AI的AI SDK。NHS的网络警报列出了应当避开的具体版本。

6月:SaaS集成。市场情报平台Klue披露,其为客户持有的OAuth令牌遭到滥用。参见案例研究2。

8月:自托管AI网关。Microsoft描述了对暴露在外、未打补丁的LiteLLM、RAGFlow和Kestra部署环境的入侵。这三款工具均已有修复版本。CISA已于6月8日将LiteLLM的漏洞列为已知被利用漏洞。

9月:评估环境。Anthropic的9月威胁报告描述了通过软件供应商获取AI密钥的尝试。云安全联盟对此做了总结。在其中1个案例中,指令被注入到某供应商的自动化评估沙箱中。

10月及11月初:开发者平台。10月,至少有9家AI开发与托管平台披露了涉及存储的客户密钥的事件。Wiz的追踪器随后列出了更多披露,总计涉及16个平台。此外还发现了被篡改的LiteLLM版本1.107.0,用户被告知回滚。

11月:共享Web供应商。据Report URI称,11月16日,一名攻击者利用Optimizely账户注册流程中的一个漏洞,将代码注入客户项目,波及178家机构的网站,其中包括HSBC、Salesforce、BBC和Ledger。该代码会替换加密货币钱包地址,并从支付表单中窃取数据;Optimizely已移除该载荷并修复了漏洞。Hudson Rock追踪到类似代码自11月13日起通过另外三个Web服务传播。这是同样的集中化模式,只是位于AI技术栈的下一层。

精选事件:日期以链接来源为准,不同来源有时会相差一天。10 月的若干条目依据的是 Wiz 和 Helix 的汇编,而非厂商的一手披露。

并不存在单一漏洞。据云安全联盟(Cloud Security Alliance)称,入侵入口各不相同。共同的目标是存储的环境变量,其中往往保存着 AI 提供商的 API 密钥。

AI 密钥具有实际价值,这解释了为何引人觊觎。CSA 援引 Anthropic 的说法指出,一个密钥可以被转售、用于算力,或被当作掩护。Wiz 估计,这些事件中被滥用的 2 个密钥,按标价计算价值超过 100,000 美元。这正是支出上限之所以重要的原因。

集中化是另一个共同因素。为众多客户存储密钥的平台自然会引来关注。CSA 将其描述为平台链模式。

没有公司能承诺自己永不被攻破。但每家公司都能控制自己如何响应。面对以机器速度行动的攻击者,这种响应依赖 4 大支柱。

1. 自动检测与遏制:以人的速度来关停已经不够。由 AI 驱动的入侵者可以在几分钟内在系统之间移动。成熟的团队会把检测直接接到遏制动作上。这意味着终止开关、自动吊销 Token 和主机隔离。Wiz 建议对批量读取密钥的行为发出告警——这是今年各起事件中最清晰的信号。最好的方案会基于该信号自动采取行动。

2. 快速且安全地沟通:速度对吊销仍然至关重要。客户需要轮换密钥,上游提供商需要在源头撤销访问权限。CSA 指出,从平台上删除密钥并不能将其吊销。成熟的团队也会谨慎选择沟通渠道。五眼联盟的一份联合公告警告不要在正被调查的网络上进行通信,而是建议采用带外通信。CISA 的操作手册同样要求在准备阶段就部署带外通信基础设施。这样,仍在窥探的攻击者就无法得知响应计划。

3. 彻底根除:堵住被利用的漏洞只是第一步。顶尖团队会假定攻击者植入了后门,并主动搜寻这些后门。他们会引入独立的取证专家,例如 Palo Alto Networks Unit 42 或 Mandiant。同一份公告也建议引入第三方事件响应支持,并警告不要只修复表象而不解决根本原因。

4. 通过缩小爆炸半径持续改进:攻击者每个月都在进步,因此安全工作永无止境。最强的主动防御是限制任何未来的攻击者所能触及的范围。零信任访问意味着任何凭证默认都不被信任。临时性架构意味着根本不保留敏感数据。数据不被存储,就无法被窃取。

一些公司已经在应用这些支柱。下面 2 个案例研究展示了具体做法:

Fireworks AI 是一个推理与训练平台。它在 2026 年 10 月披露了一起事件,而那个月对该行业而言异常繁忙。它的响应是一个有力范例,展示了围绕这 4 大支柱、成熟且结构稳健的事件响应。

支柱 1:检测与遏制

Fireworks 于 10 月 6 日发现内部凭据被未经授权使用。第二起事件始于 10 月 24 日,经由托管其研发机器的第三方云服务商发生。Fireworks 检测到了该活动,并在其开始当天就将其关闭。这正是机器速度的威胁所要求的运营敏捷性。

第二支柱:快速通报

Fireworks 于 10 月 7 日开始通知受影响客户,距发现约 1 天。它于 10 月 9 日发布了公开公告,并随着调查发现不断更新。它还通知了执法部门。公告列出了受影响的资源元数据和账单信息。对某些客户,公告还列出了数据集、不完整的模型权重和密钥。

第三支柱:彻底清除

Fireworks 聘请 Palo Alto Networks 的 Unit 42 进行取证分析。引入独立的顶级团队是一个成熟的决定。这意味着调查不止于最初的入口点,还会排查攻击者可能留下的任何东西。Fireworks 还在 10 月 9 日前完成了员工 API 密钥的轮换。

第四支柱:以设计缩小影响范围

Fireworks 最重要的防护措施此前就已到位。它表示未发现其 Inference、Training SDK 和 Nexus 产品受到影响。这 3 款产品默认以零数据留存运行,因此提示词和输出不会静态存储。这一架构使核心产品免受此次所报告的影响。此后 Fireworks 更进一步:它暂停了 Batch API,其文档称已存储的 Batch 输入和输出将被永久删除。企业客户还可以在整个账户层面强制执行 ZDR 策略。

2026 年 11 月 13 日,Fireworks CEO Lin Qiao 在 X 上发帖解释公司的数据留存政策。她描述了 2 种标准 API 设计:

她将其比作把电子表格保存在云盘里,与把其中的数字交给计算器运算之间的区别。在这两种设计中,客户都掌控着自己的数据。她的结论是:只要工作流程允许,临时处理更为可取。它不会留下被遗忘的数据集,也不会出现遗漏的删除调用,而每一份被保留的副本都是需要保护的对象。在我们看来,这是主动防御的黄金标准,而 10 月发生的一系列事件也印证了这一点。

包括 Cursor、Uber 和 Notion 在内的一些客户已发布了自己的通知。调查仍在进行中。

教训是:当天遏制、快速通知以及独立取证,能够缩短事件持续的时长。而临时性架构则从根源上限制了可被触及的范围。

原文较长,译文为节选;译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(Not If, But How: Responding to Security Incidents in the Age of AI Attackers)

Attackers now have powerful AI on their side. AI agents can scan, plan and run intrusions at machine speed, and they improve every month. The result is more attacks, more often and at global scale. In that environment, more companies will face a security incident at some point. What separates them is how they respond. Machine-speed threats call for a modern playbook with 4 pillars. Detect and contain automatically. Communicate fast and securely. Eradicate thoroughly. Keep shrinking the blast radius. This article shows the trend through 6 months of incidents. It then looks at a few companies, with Fireworks AI as a model of mature, structurally sound response.

TL;DR

  • The trend: AI-powered attackers are driving more frequent attacks at global scale.
  • The evidence: Several AI platforms disclosed security incidents between May and November 2026.
  • Pillar 1: Detect and contain automatically. Human-speed shutdowns are too slow.
  • Pillar 2: Communicate fast and securely, over out-of-band channels.
  • Pillar 3: Eradicate thoroughly, with outside forensic experts.
  • Pillar 4: Keep shrinking the blast radius. Data that is never stored cannot be stolen.
  • Case studies: Fireworks AI and Klue show these pillars in practice.

Why are attacks becoming more frequent?

AI has changed the economics of attacking. Tasks that took skilled people days can now be delegated to agents.

In November, Google Threat Intelligence Group and Mandiant said AI misuse had reached a “critical inflection point”. They report that attackers use AI to shorten timelines and run multi-stage intrusions at machine speed. CrowdStrike reported that 1 group used AI agents against more than 200 organizations in 11 days. Palo Alto Networks’ Unit 42 found attacks moving about 4 times faster than a year earlier.

The shift started earlier. In 2025, Anthropic disrupted a campaign where AI did 80 to 90% of tactical work with little human input.

The tools keep improving, so the pace will keep rising. AI infrastructure is an especially attractive target, because AI access itself has value. The last 6 months show what that looks like.

What was disclosed in the last 6 months?

Incidents appeared across every layer of the AI stack. Together they show how broad the activity has become.

May: package registries. On May 10 and 11, the “Mini Shai-Hulud” campaign altered about 170 npm packages and some PyPI libraries. Affected projects included AI SDKs from Mistral AI and Guardrails AI. The NHS cyber alert lists the specific versions to avoid.

June: SaaS integrations. Klue, a market intelligence platform, disclosed misuse of OAuth tokens it held for customers. See case study 2.

August: self-hosted AI gateways. Microsoft described intrusions into exposed, unpatched LiteLLM, RAGFlow and Kestra deployments. Fixed versions exist for all 3 tools. CISA had listed the LiteLLM flaw as known exploited on June 8.

September: evaluation environments. Anthropic’s September threat report described attempts to obtain AI keys through software vendors. The Cloud Security Alliance summarized it. In 1 case, instructions were injected into a vendor’s automated evaluation sandbox.

October and early November: developer platforms. At least 9 AI developer and hosting platforms disclosed incidents involving stored customer secrets in October. Wiz’s tracker later listed further disclosures, for 16 platforms in total. A tampered LiteLLM release, version 1.107.0, was also identified, and users were told to roll back.

November: shared web vendors. On November 16, an attacker abused a flaw in Optimizely’s account signup flow to inject code into customer projects, reaching 178 organizations’ websites, according to Report URI, including HSBC, Salesforce, the BBC and Ledger. The code swapped crypto wallet addresses and skimmed payment forms; Optimizely removed the payload and fixed the flaw. Hudson Rock traced similar code through three other web services, beginning November 13. It is the same concentration pattern, one layer below the AI stack.

Date (2026)PlatformWhat was affectedEntry point, as reported
May 10 to 11npm and PyPI packages, incl. Mistral AI and Guardrails AI SDKsDeveloper and CI credentials on installing hostsPoisoned package releases
June 11 to 12KlueCustomer Salesforce and Gong dataLegacy credential, then stored OAuth tokens
AugustLiteLLM, RAGFlow, Kestra deploymentsModel provider keys, backend accessExposed, unpatched gateways
October 6Fireworks AIEnvironment variables in evaluation jobs, limited accountsMisused internal credentials
October 14 to 15Zeabur3,218,965 environment variable records, 321,810 accountsLong-lived admin tokens, exposed admin interface
October 15Lightning AICredentials; Studio taken offlineCredential access
October 16 to 17OpenCodeOpenAI, Anthropic and Google keys of about 1.4% of Console usersLegacy system
October 20 to 21RailwayEncrypted environment variables, some decryptedPrivilege escalation in build infrastructure
October 21LiteLLM 1.107.0Cloud credentials and LLM keys on installing hostsRelease pipeline compromise
October 23WarpAgent decision logs and telemetry, about 1% of usersThird-party vendor dependency
October 23 to 24Fireworks AIBilling data; for some customers, datasets and incomplete model weightsIncident at a third-party cloud provider
October 29CursorAltered extension copy in its marketplace cache, up to 5,000 usersCredential from exposed internal data
October 30NorthflankCustomers told to assume environment variables were decryptedNot disclosed
November 8NEAR AIInference server; 2 API keys exposed in logsRemote code execution flaw in SGLang

Selected incidents: Dates follow the linked sources, which sometimes differ by a day. Several October entries rely on compilations by Wiz and Helix rather than primary vendor disclosures.

What do these incidents have in common?

There was no single vulnerability. According to the Cloud Security Alliance, entry points varied. The common target was stored environment variables, often holding AI provider API keys.

AI keys carry real value, which explains the interest. CSA, citing Anthropic, notes that a key can be resold, used for compute or used as cover. Wiz estimated that misuse of 2 keys from these incidents exceeded $100,000 at list price. That is why spending caps matter.

Concentration is the other shared factor. A platform that stores secrets for many customers naturally draws attention. CSA describes this as the platform-chain pattern.

What does a modern incident response playbook look like?

No company can promise it will never be breached. Every company can control how it responds. Against attackers that move at machine speed, that response rests on 4 pillars.

1. Detect and contain automatically: A human-speed shutdown is no longer enough. An AI-driven intruder can move between systems in minutes. Mature teams wire detection directly to containment. That means kill-switches, automatic token revocation and host isolation. Wiz recommends alerting on bulk secret reads, the clearest signal in this year’s incidents. The best setups act on that signal automatically.

2. Communicate fast, and securely: Speed still matters most for revocation. Customers need to rotate keys, and upstream providers need to revoke access at the source. CSA notes that deleting a key from a platform does not revoke it. Mature teams also choose the channel with care. A Five Eyes joint advisory warns against communicating over the network under investigation. It advises out-of-band communication instead. CISA’s playbooks likewise call for out-of-band communications infrastructure as part of preparation. That keeps an attacker who is still watching from learning the response plan.

3. Eradicate thoroughly: Closing the hole that was used is only the first step. Top teams assume the attacker planted backdoors and hunt for them. They bring in independent forensic experts, such as Palo Alto Networks Unit 42 or Mandiant. The same advisory recommends third-party incident response support. It also warns against fixing symptoms instead of the root cause.

4. Keep improving by shrinking the blast radius: Attackers improve every month, so security is never finished. The strongest proactive defense limits what any future attacker can reach. Zero Trust access means no credential is trusted by default. Ephemeral architecture means sensitive data is not kept at all. If data isn’t stored, it can’t be stolen.

Some companies already apply these pillars. The 2 case studies below show how:

Case study 1: How did Fireworks AI respond?

Fireworks AI is an inference and training platform. It disclosed an incident in October, 2026, during a busy month for the industry. Its response is a strong example of mature, structurally sound incident response across the 4 pillars.

Pillar 1: detect and contain

Fireworks identified unauthorized use of internal credentials on October 6. A second event began on October 24, through a third-party cloud provider that hosted its R&D machines. Fireworks detected that activity and closed it on the same day it began. That is the kind of operational agility machine-speed threats demand.

Pillar 2: communicate fast 

Fireworks began notifying affected customers on October 7, about 1 day after detection. It published a public bulletin on October 9 and has updated it as findings came in. It also notified law enforcement. The bulletin lists affected resource metadata and billing information. For certain customers, it also lists datasets, incomplete model weights and secrets.

Pillar 3: eradicate thoroughly 

Fireworks engaged Palo Alto Networks Unit 42 for forensic analysis. Bringing in an independent, top-tier team is a mature decision. It means the investigation looks beyond the first entry point, for anything an attacker may have left behind. Fireworks also completed rotation of employee API keys by October 9.

Pillar 4: a smaller blast radius by design

Fireworks’ most important protection was already in place. It says it detected no impact to its Inference, Training SDK and Nexus products. All 3 run with zero data retention by default, so prompts and outputs are not stored at rest. That architecture kept the core products out of the reported impact. Fireworks has since gone further. It paused its Batch API, and its documentation says stored Batch inputs and outputs will be permanently deleted. Enterprise customers can also enforce an account-wide ZDR policy.

The philosophy behind it: 

On November 13, 2026, Fireworks CEO Lin Qiao posted on X to explain the company’s data retention policy. She described 2 standard API designs:

  • Customer-managed retention (upload, use, delete). A user uploads a dataset, runs jobs on it, then deletes it. Batch APIs are the typical example. The platform keeps the data until told to delete it, because it cannot know when the user is finished. Cleanup is the user’s job. Fireworks says it does not retain data after deletion.
  • Ephemeral processing (send, compute, return). The platform processes a request and returns the result without keeping the content. This is zero data retention. Fireworks applies it to its inference APIs, its Training SDK, including serverless training, and Nexus. Prompts and generations are not persisted at rest.

She compared it to saving a spreadsheet in a cloud drive versus passing its numbers to a calculator. In both designs, customers control their data. Her conclusion is that ephemeral processing is preferable whenever the workflow allows it. It leaves no forgotten datasets or missed deletion calls, and every retained copy is something to protect. In our view, this is the gold standard for proactive defense, and the October events bear it out.

Some customers, including Cursor, Uber and Notion, have published their own notices. Investigations are ongoing.

The lesson: same-day containment, fast notice and independent forensics limit an incident’s duration. Ephemeral architecture limits what can be reached at all.

Case study 2: How did Klue respond?

Klue is a market intelligence platform that connects to customer CRM and sales tools. Its June incident shows the same habits in a different part of the stack.

What happened. On June 11 and 12, an attacker used a compromised legacy credential tied to Klue’s integration service. With it, they obtained OAuth tokens Klue held for customers. Those tokens were used to read data from some connected Salesforce and Gong environments.

Pillars 1 and 2: detect and communicate. Klue identified the activity on June 12 and notified customers the same day. Same-day notice let customers start reviewing their own connected systems at once.

Pillars 1 and 3: contain and eradicate. Klue revoked affected credentials and disabled 9 integrations, including Salesforce, HubSpot, Gong and Slack. Salesforce disabled the Klue integration on June 17.

The lesson: every stored token is standing access. Retiring old credentials and long-lived tokens is pillar 4 in practice for integration platforms.

What do the two case studies have in common?

Both companies held something on behalf of customers: datasets and secrets at Fireworks, OAuth tokens at Klue. Both detected the activity and notified customers within about a day. Both cut off access at the source. Fireworks went further on pillars 3 and 4, with independent forensics and an architecture that stores less by default.

Who is responsible when an AI platform has an incident?

Usually more than 1 party. Most of these events ran through a chain of companies, and each link plays a role.

Upstream suppliers: Cloud providers, package registries and build tools sit beneath every platform. Fireworks’ second event began at a third-party cloud provider. The May package campaign and the LiteLLM release began in build and publishing pipelines.

The platform or provider: A platform that holds customer secrets is responsible for protecting them. CSA calls platform operators a main point of control.

The vendor building on the platform: Companies that integrate an AI API decide what they send, store and share. When they use upload-based features, deleting data they no longer need is part of their job. In Cursor’s case, the key opened a bucket Cursor used to share data with Fireworks. Some traces in it held unredacted credentials, which reached Cursor’s own tools. Cursor has since added secret scanning and rotated its credentials.

The end customer: Keys stay valid until the issuing provider revokes them. CSA suggests treating each long-lived key as exposed if a platform holding it reports an incident.

Good choices at any link can contain an incident. That is why the guidance below is split by role.

What do the incidents show about stored data?

A useful lesson comes from what was not affected. Cursor reports that prompts and code from Privacy Mode users were not affected. That data was never kept, so there was nothing to copy.

The pattern holds across the timeline. Incidents involved stored items: environment variables, OAuth tokens, datasets, logs and traces. Live requests processed in memory were not among the reported impacts.

This is the practical case for ephemeral processing. Fewer stored copies mean fewer things to protect, track and delete.

What should each party do next?

The guidance below draws on CSA, Wiz and government incident response advice.

AI providers and platforms

  1. Deploy automated containment triggers for bulk secret reads. Revoke tokens and isolate workloads without waiting for a human.
  2. Retire non-expiring admin tokens, and keep admin interfaces off the public internet.
  3. Separate internal workloads from customer workloads.
  4. Set up out-of-band communication channels before an incident, and use them during it.
  5. Notify customers, partners, upstream providers and law enforcement quickly. Measure notice time in hours.
  6. Retain an independent forensic firm in advance, so eradication can start on day 1.
  7. Adopt ephemeral architecture and zero data retention wherever workflows allow. Offer spending caps and short-lived credentials by default.

Vendors building on AI APIs

  1. Prefer ephemeral, zero data retention endpoints. As Lin Qiao puts it, use them whenever the workflow allows.
  2. When you do upload data, own the delete step. Set expiry dates on shared buckets and datasets, and delete them when the job ends.
  3. Scan prompts, traces and datasets for secrets before they leave your systems.
  4. Issue 1 key per platform and per application, so revocation stays narrow.
  5. Ask suppliers how fast they detect, contain and notify, and whether they use outside forensics.

Every team holding AI keys

  1. Rotate at the issuing provider. Deleting a key from a platform does not revoke it.
  2. Set spending and rate caps that match expected use.
  3. Review usage and billing logs after any supplier notice.

Key Takeaways

  • AI-powered attackers are making attacks faster, more frequent and more global.
  • Modern response rests on 4 pillars: contain automatically, communicate securely, eradicate thoroughly, shrink the blast radius.
  • Zero data retention kept companies like Fireworks’ core products out of the reported impact.
  • If data isn’t stored, it can’t be stolen.

Further reading: Wiz, Black October for AI developer platforms, CSA, Black October research note and CSA, platform-chain campaigns.


Sources


Thanks to the Fireworks team for the thought leadership / resources for this article. This article is sponsored.

The post Not If, But How: Responding to Security Incidents in the Age of AI Attackers appeared first on MarkTechPost.

正文保留英文原文,标题与摘要由机器翻译。

本文由采集器自动同步

原文发布于 MarkTechPost(Not If, But How: Responding to Security Incidents in the Age of AI Attackers),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗