← 返回列表

AI Infrastructure Keeps Getting Attacked — NEAR AI’s SGLang Breach Is Just the Latest Warning

高级反反爬 Security Boulevard 发布 2026-11-18 入库 2026-11-19T04:51:08

Summary

Author: Koby Bar, Senior Vice President of Product

On November 9, 2026, NEAR AI disclosed that an attacker had exploited a vulnerability in SGLang, an open-source LLM inference engine, exposing account details—including names and email addresses—for roughly 100 users, and chat histories, settings, and logs from 18 legacy AI agents that weren’t running in its confidential-computing environment.

It’s the latest reminder that AI infrastructure is where attackers are now focusing. Langflow, a widely used open-source platform for building AI agents and retrieval-augmented generation (RAG) pipelines, has served as a similar proof point over the last year. It has racked up at least seven critical CVEs since mid-2025, including CVE-2026-33017, which was disclosed in March 2026 as an unauthenticated remote code execution (RCE) flaw in a public Langflow endpoint that was unauthenticated by design. It was weaponized within 20 hours of public disclosure—before a proof-of-concept exploit was even circulating publicly, according to researchers at Sysdig.

Both SGLang and Langflow are examples of the kinds of platforms and tools that have quickly appeared inside enterprise environments as teams race to stand up AI capabilities. Together, they highlight the same risk pattern: security programs haven’t caught up to how fast AI infrastructure is being deployed.

Why AI Infrastructure Keeps Getting Attacked

It’s tempting to read NEAR AI’s SGLang breach as an isolated incident, or Langflow’s CVE count as a Langflow-specific problem. But it’s more useful to read both as previews of what security leaders can expect to keep happening across AI infrastructure broadly, and a guide to how security programs need to evolve to combat this new reality.

Inference servers like SGLang sit on some of the most sensitive assets in an AI stack: API keys, live request data, and session context. That concentration of risk is exactly what played out at NEAR AI—two customers had API keys exposed in server logs, and the attacker could see live requests being processed on the compromised server.

Agent-orchestration platforms like Langflow carry a similar concentration of risk. Langflow’s Global Variables store is designed to hold things like model API keys, cloud credentials, and database passwords, all in one place so AI workflows can run smoothly. While that is a sensible design for an AI orchestration tool, it also creates the kind of concentrated, high-value target that security teams historically haven’t had on their radar because it didn’t exist in their environment two years ago.

Enterprises are standing up AI agents, RAG pipelines, and model context protocol (MCP)-based tooling faster than security programs are extending coverage to support them. Most vulnerability scanning, patch cadences, and exposure management programs were built around networks, endpoints, cloud workloads, and identity—not around the orchestration layer that now sits between your data and a model that can act on it.

Attackers aren’t waiting for security teams to catch up. IBM’s 2026 Cost of a Data Breach Report found that AI-enabled breaches now average $6 million, roughly $1 million above the global average, and that more than one in five organizations have already had a breach targeting their own AI models or applications. Governments are sounding the same alarm: a June 2026 joint statement from the Five Eyes cyber agencies warned boards and executives directly that “AI is shortening the time between vulnerability discovery and exploitation,” and that “cyber risk assumptions can become outdated in months, not years.”

The Question Isn’t Whether You’re Patched. It’s Whether You’re Exposed.

Patch compliance and real-world exploitability are not the same thing—and that gap is where most AI infrastructure risk actually lives. For example, confirming that Langflow is on the latest version tells you a vulnerability has been remediated. It doesn’t tell you whether your specific deployment—your Global Variables configuration, your network exposure, your guardrails, your connected credentials—was ever actually reachable by the techniques attackers used against CVE-2026-33017. It also doesn’t tell you what else is exposed right now that hasn’t made it into a CVE yet.

NEAR AI’s own breach report makes a related point. Its confidential-computing hardware worked exactly as designed, protecting data from the operator and the host. What it didn’t do—because it isn’t built to—was protect against a vulnerability in the software running inside that environment, and that’s exactly where the attacker got in. The control held up; the assumption about what it covered didn’t, and nobody had tested that assumption against a real attack. Having a control in place isn’t the same as knowing it will hold during an actual incident.

That’s exactly what adversarial exposure validation (AEV) is built to answer—and why AI infrastructure belongs inside a continuous threat exposure management (CTEM) program rather than a one-off “AI security” initiative. CTEM is built on the premise that your exposure changes constantly—new deployments, new configurations, new attacker techniques—and that the only way to keep pace is to validate continuously, not to scan on a quarterly schedule and hope nothing changes in between. NEAR AI’s SGLang breach and Langflow’s CVE count are both textbook cases for why that premise holds.

AI Infrastructure Exposure Validation: Testing the Way Attackers Do

Most security programs are not yet testing AI infrastructure with the same attacker-eye-view rigor applied to everything else in their environment.

SafeBreach AI Security Validation is content within the SafeBreach CTEM Platform that safely simulates real-world attacks against an organization’s AI applications and AI infrastructure to prove whether its defenses stop them. Simulations are run against your production model with your guardrails, system instructions, and knowledge base in place, and each returns a clear Blocked or Not Blocked verdict and remediation guidance.

Results map to frameworks security teams already report against, including the OWASP Top 10 for LLM Applications, the OWASP Top 10 for Agentic Applications, and MITRE ATLAS.

SafeBreach AI Security Validation covers four critical areas:

  1. Guardrails Validation: Tests whether AI guardrails stop data leakage, system prompt leakage, harmful content generation, and malicious code generation. Supports models on Amazon Bedrock and Azure OpenAI.
  2. AI Infrastructure Testing: Tests whether attackers can exploit known vulnerabilities in AI orchestration software.
  3. Shadow AI Validation: Tests whether data loss prevention (DLP) controls stop employees from sharing sensitive data with public AI chatbots.
  4. Custom AI Attacks: Lets teams build their own AI attacks with the SafeBreach AI Security skill in VS Code.

This kind of continuous validation is built to keep pace with AI-driven threats. It isn’t a one-time AI risk assessment or a patch tracker. It’s ongoing proof of whether the systems holding your model keys, cloud credentials, and session data would stand up to the techniques attackers are using against AI infrastructure today.

Don’t Wait for the Next Breach to Prove the Point

AI infrastructure will continue to be the focus of malicious actors, and the underlying dynamic isn’t specific to one platform or tool. The risk is applicable to AI infrastructure broadly and is a symptom of the fact that it has been adopted faster than testing can keep up at most organizations.

Schedule a demo to see how SafeBreach’s CTEM Platform and AI Security Validation content test AI infrastructure alongside the rest of your environment—continuously, not just when the next CVE makes headlines.

About the Author

Koby Bar is Senior Vice President of Product at SafeBreach, where he leads the company’s product strategy and roadmap. With more than 16 years of experience in product leadership across early-stage startups and large, established organizations, Koby has a strong track record of turning complex security challenges into products customers rely on. He leads a team of product managers, designers, and technical writers focused on delivering an exceptional user experience and maximizing the value of the SafeBreach CTEM Platform. Most recently, he has driven the company’s AI strategy, including the launch of SafeBreach Helm, the LLM-driven engine that connects the pillars of CTEM, and the company’s new AI agents. Before joining SafeBreach in 2022, Koby held product leadership roles at IRONSCALES and Rogers Communications. He is a University of Toronto alumni.

原文链接:https://securityboulevard.com/2026/11/ai-infrastructure-keeps-getting-attacked-near-ais-sglang-breach-is-just-the-latest-warning/