Why the SGLang Breach Changes AI Vendor Risk
An attacker reached inside the one boundary NEAR AI told customers would hold, and the boundary held exactly as designed. That is the uncomfortable part of NEAR AI's own account of its November 2026 security incident. The company did not claim its confidential computing environment failed. It said the environment worked. The vulnerability sat in the software running inside it, and that software had never been tested the way the hardware boundary around it had been.
Enterprises evaluating AI infrastructure vendors have learned to ask one question above most others, whether this platform runs sensitive workloads inside a confidential computing environment or a trusted execution environment, sometimes called a secure enclave. A yes has become something close to a pass on vendor risk review. The NEAR AI breach is a disclosed, dated, named case where that yes was true and the data still got out, because the question itself was incomplete.
What leaked was not incidental telemetry. It was account details for roughly 100 users, and chat history, settings and logs belonging to 18 AI agents that ran on a legacy deployment outside the confidential computing environment, plus API keys belonging to two Cloud customers, exposed in server logs the attacker could read. Those are the categories of data a vendor's isolation promise exists to protect. When a single remote code execution flaw in an open source inference engine put all three in reach, the promise did not fail quietly. NEAR AI published the details itself.
This piece is written for the reader who signs off on that vendor risk review, usually a CISO or a chief compliance officer who must tell a board, a regulator or an auditor why a third-party AI platform was trusted with sensitive data. The argument is not that confidential computing is worthless. It is that a vendor's isolation architecture is a claim about a moment in time, and the moment a new vulnerability is disclosed in anything running inside that architecture, the claim needs to be checked again, not assumed to still hold.
Key Takeaways
- Confidential computing protects against the operator, not against software bugs. NEAR AI's own incident report says its isolation hardware worked as intended. The attacker got in through a remote code execution flaw in the SGLang inference engine running inside that environment, not through a failure of the enclave itself.
- The exposed data was exactly what the boundary was supposed to contain. Account details for about 100 users, data from 18 AI agents that sat outside the confidential computing environment, and API keys for two Cloud customers all moved from inside NEAR AI's infrastructure to an outside attacker.
- The gap between disclosure and exploitation is now measured in hours. A separate, unrelated Langflow flaw tracked as CVE-2026-33017 was weaponized about 20 hours after its public advisory, before any proof-of-concept exploit code was circulating.
- AI infrastructure vulnerabilities are piling up faster than vendors can patch them. Security researchers count at least seven critical Langflow flaws disclosed since mid-2025, with five of them now sitting on CISA's Known Exploited Vulnerabilities catalog.
- A one-time vendor attestation is already outdated the day after it is signed. Continuous, verifiable enforcement over what an AI agent or inference platform can touch must replace a point-in-time audit of its architecture, because the architecture is never the whole story.
What NEAR AI's Own Report Admits
NEAR AI published its own incident report on November 9, 2026, describing a remote code execution vulnerability in SGLang, an open source large language model inference engine running on its servers. An attacker used that flaw to gain code execution on one inference server and from there reached an internal system behind a deprecated feature called Private Chat. The company's own words matter here, because they are more precise than most vendor breach disclosures get. NEAR AI wrote that confidential computing protects data from the operator and the host, and that it does not protect against a vulnerability in the software running inside the environment, which is what happened.
The scope NEAR AI confirmed is specific. Account details, meaning email addresses, names and in some cases plan and balance information, were downloaded for roughly 100 users. Chat history, settings and logs were accessed for 18 agents running on a legacy NEAR AI Agents deployment, a deployment the company states plainly was not running inside its confidential computing environment. Server logs exposed API keys belonging to two Cloud customers, and those keys have since been revoked. NEAR AI also said the attacker could have observed live requests being processed on the compromised server on November 8, including requests submitted in confidential mode, though the company reports no evidence that stored Private Chat conversations or uploaded files were read.
The Block coverage of the disclosure adds useful context without contradicting NEAR AI's own numbers, confirming the same user and agent counts and noting that Cloud access was paused during the investigation and restored once remediation was complete. NEAR AI's remediation list reads like a standard incident response runbook. It terminated active sessions and admin credentials, rotated agent tokens, patched the vulnerability across all inference servers, reported it upstream to the SGLang project, completed a forensic review, and notified affected users and agent owners individually. None of that is a criticism of NEAR AI's response, which was fast and unusually transparent for an AI infrastructure vendor. The point is narrower. Even a vendor that does most things right after a breach cannot undo the fact that, for a window of time, data inside its confidential computing boundary was reachable from outside it.
The Isolation Promise Enterprises Are Buying
Enterprises have good reason to want a yes-or-no answer to the question "does this vendor's infrastructure run inside a trusted execution environment." It is concrete, it is auditable, and it maps neatly onto a procurement checklist. A confidential computing claim tells a buyer something real. The cloud provider that physically hosts the workload, and the vendor's own system administrators, cannot read the data processed inside the enclave even if they wanted to. For workloads involving regulated data, that is a meaningful control, and it is one reason AI inference and agent platforms market it so heavily.
The trouble is that this specific guarantee answers a narrower question than the one most buyers think they are asking. A buyer's real question is usually closer to "can anyone outside my organization reach my data through this vendor." Confidential computing answers "can the vendor's own staff and its cloud host reach your data," which is a subset of the real question, not the whole of it. Vendor data breaches have become the top legal liability vector for exactly this reason. The contractual and technical controls a buyer reviewed at onboarding rarely cover every path an attacker can take once they are inside the vendor's own perimeter, enclave or not.
NEAR AI's disclosure is useful precisely because it names the gap instead of leaving it implicit. The legacy AI agents that were compromised sat outside the confidential computing environment entirely, which the company has now acknowledged and says it is addressing by extending the same security rigor to its surrounding services. But the inference server that let the attacker in was inside the isolation boundary, running software, SGLang, that had not received the same scrutiny as the enclave hardware around it. A buyer who asked only "is this inside a TEE" during due diligence would have gotten a technically accurate yes and still ended up exposed.
Why the Enclave Held and the Software Did Not
It is worth being precise about what happened, because the distinction between hardware isolation and software security is where most vendor risk assessments go wrong. A trusted execution environment is a hardware-backed guarantee about who can read memory and data inside a defined boundary. It does not inspect, patch or secure the application code running inside that boundary. If that code has a remote code execution vulnerability, as SGLang did, an attacker who exploits it gains the same level of access to data inside the enclave that the legitimate application has. The enclave does not stop the attacker from reading what the application itself is already permitted to read.
MCP server security runs into the same category error. Buyers ask whether the connection is encrypted, whether access is authenticated, whether the deployment follows a zero trust architecture. All reasonable questions, and none of them address whether the orchestration or inference software sitting between the agent and the data has its own exploitable bugs. A control at the transport layer or the hardware layer cannot catch a flaw in the application layer, and increasingly the application layer is where AI infrastructure breaks.
This is also why the distinction between NEAR AI's confidential computing environment and its legacy agents deployment matters so much. The agents outside the enclave were exposed through the same intrusion, by the same attacker, in the same incident, just without even the partial protection the enclave offered the inference server. A security architecture with two tiers, one hardened and one legacy, is common in real enterprise environments, including a buyer's own. A CISO should ask a vendor a harder follow-up question than what protects the newest, best-architected part of your stack. That follow-up is what protects everything else, and how do you know.
The Window Between Disclosure and Exploitation
A second data point from the same week sharpens the argument. Sysdig's research on CVE-2026-33017, an unauthenticated remote code execution flaw in Langflow, an open source framework for building AI agent and retrieval-augmented generation pipelines, found the first exploitation attempt about 20 hours after the public advisory went out, before any proof-of-concept code was circulating publicly. Attackers built a working exploit directly from the advisory's description of the flaw. Sysdig's honeypot fleet recorded automated scanning, credential harvesting and environment variable dumps from six distinct source IPs within 48 hours of disclosure.
Twenty hours is not an outlier anymore. A SafeBreach analysis published alongside its coverage of the NEAR AI breach counted at least seven critical Langflow vulnerabilities disclosed since mid-2025, with five of them now listed on CISA's Known Exploited Vulnerabilities catalog, meaning the federal government has confirmed active exploitation in the wild. One of those additions, documented in CISA's own advisory covering CVE-2026-33017 specifically, set a remediation deadline of only a couple of weeks for federal agencies, a timeline that assumes organizations are already watching for new entries rather than reviewing their AI vendor stack once a year.
Put the NEAR AI incident and the Langflow pattern next to each other and a shape emerges that neither story shows on its own. AI infrastructure, whether it is an inference engine like SGLang or an agent-building framework like Langflow, is accumulating critical vulnerabilities at a pace that outstrips any reasonable patch cycle, and the time attackers need to weaponize a new one has collapsed to hours. A vendor risk review that asks about architecture once, at onboarding, is reviewing a snapshot of a system that may have a dozen new critical flaws by the time the contract renews.
What the Breach Data Shows
Some of this risk can be quantified, and the numbers argue against treating AI infrastructure breaches as rare events. IBM's breach report found that more than 20 percent of surveyed organizations reported a breach targeting their AI models or applications specifically, not their traditional IT estate. The same report found that one in four malicious breaches are now AI-enabled, a 56 percent increase, and that these AI-enabled breaches cost organizations an average of $6 million, against a global average breach cost of $4.99 million. IBM's own newsroom release attributes that figure directly to the report, dated July 29, 2026.
These numbers describe the cost of a breach once it has happened, which is a different and narrower question than whether an enterprise can detect the exposure fast enough to limit it. NEAR AI's timeline, by its own account, ran from an exploited vulnerability on November 8 to a published, detailed incident report on November 9, a turnaround most enterprises cannot match for their own environments, let alone a vendor's. That speed came from NEAR AI's own forensic capability, not from any control the buyer had purchased or configured. A buyer relying entirely on a vendor's own goodwill and competence to discover and disclose a breach has outsourced the most important part of its own risk posture.
From One-Time Attestation to Continuous Enforcement
The honest conclusion is not that confidential computing is a false promise. It is that an architecture claim, verified once, answers a question that keeps changing. The trust boundary covers the same dynamic from a different angle. Every new connector, every new piece of infrastructure a vendor stands up to serve an AI agent, moves the edge of what an enterprise is implicitly trusting, often without the enterprise ever reviewing the change. A confidential computing attestation signed at onboarding says nothing about the inference engine version the vendor swapped in six months later, or the legacy deployment the vendor never migrated into the hardened environment at all.
What a defensible posture requires is enforcement that does not depend on trusting the vendor's architecture to stay static. Contextual Data Enforcement applies that logic at the point where an AI agent or inference platform is about to touch enterprise data, evaluating the request against the entities, relationships and context involved rather than relying on a permission granted once and assumed to still be appropriate. The same principle extends to how an organization connects agents to infrastructure in the first place. Agentic data security sits at exactly the chokepoint where a vendor's internal architecture stops being the enterprise's problem and the data flowing through that connector starts being the enterprise's problem again, regardless of what the vendor's enclave does or does not contain.
MCP Python SDK flaw coverage made a closely related point earlier this year. A single authentication bug in widely used AI agent infrastructure exposed how much enterprises trust a protocol's reputation rather than its continuously verified behavior. AI governance holds up only when a trust boundary is retested, since one drawn once and never checked again is not a boundary an enterprise can defend to a regulator after the fact. What changes the conversation is enforcement that runs continuously, at execution time, rather than a review that runs once, at signing time.
What CISOs and Compliance Leads Should Ask Vendors Now
A revised vendor risk conversation must move past the single confidential computing question. It should ask which components of the vendor's stack sit inside the hardened environment and which, like NEAR AI's legacy agents, sit outside it on older infrastructure. It should ask how quickly the vendor patches disclosed vulnerabilities in the software running inside its enclave, beyond just the enclave's own configuration. It should ask whether the vendor tests its own dependencies against the kind of rapid weaponization Sysdig documented, or whether patching happens on a calendar cycle that assumes attackers move slower than they do. An AI risk assessment that stops at architecture review is answering last year's question.
None of this removes the enterprise's own responsibility to verify what it can independently confirm. Non-human identity governance covers why credentials issued to an AI agent or an API integration need the same scoped, revocable, individually auditable treatment as a human user's login, a lesson NEAR AI's exposed API keys underline directly. Static API keys that never rotate and carry broad scope turn a single server compromise into a multi-customer incident, which is close to what happened with the two Cloud customers whose keys sat exposed in NEAR AI's logs. An enterprise that scopes and rotates its own credentials aggressively limits the blast radius of a vendor breach regardless of what that vendor's architecture promised.
An exploit window this narrow means a vendor review conducted on an annual cycle is reviewing a system that no longer exists by the time the review is filed. An enterprise cannot patch a vendor's software directly. It can insist on continuous disclosure commitments, scope its own exposure through its own credential hygiene, and build enforcement at the data layer that does not depend on any single vendor's isolation claim remaining accurate indefinitely. That combination, not a stronger checkbox on a procurement form, is what survives the next disclosure.
Frequently Asked Questions
1. We're evaluating an AI inference vendor that advertises confidential computing and trusted execution environments. What does that protection cover, and what does it leave exposed?
Confidential computing and trusted execution environments protect data from the cloud operator and the underlying host hardware, isolating a workload from the infrastructure around it. They do not protect against a vulnerability in the application software running inside that isolated environment. NEAR AI's November 2026 breach confirmed this directly, an attacker exploited a flaw in the SGLang inference engine running inside its confidential computing environment and reached data that boundary was meant to contain.
2. As the compliance lead at a healthcare company sending patient records through a third-party AI platform, how do we keep checking that a vendor's isolation claims still hold after our initial security review is done?
A healthcare compliance team should treat a vendor's isolation claim as a starting point, not a permanent fact, by requesting updated vulnerability disclosures on a recurring schedule, confirming patch timelines for the software running inside the isolated environment, and asking whether the vendor tests that environment's own dependencies on an ongoing basis rather than only at onboarding. Annual attestations alone will not catch a new flaw disclosed between reviews.
3. Our AI infrastructure vendor just disclosed a breach of a system it said was isolated. Should we pause sending it sensitive data, or is it safe to continue?
Whether to continue sending sensitive data should depend on what the vendor's disclosure says was exposed, patched and verified, not on the fact that a disclosure happened. NEAR AI reported it patched the exploited vulnerability, rotated affected credentials and restored Cloud access once its investigation was complete, and attributed the exposure to application software rather than a failure of its isolation hardware. A buyer should independently verify that timeline before assuming the risk is closed.
4. If our AI inference vendor is breached through a system meant to keep our data isolated, what will an auditor expect us to produce about detection time and data exposure?
An auditor will expect a precise timeline covering first unauthorized access, detection, containment and the patch, alongside an exact account of which data categories were exposed. NEAR AI's own incident report supplies dates for the exploitation and disclosure, names the specific data categories involved, and states what evidence shows was not accessed. A buyer should be able to produce that same vendor-supplied timeline on request rather than reconstruct it after an auditor asks.
5. What is the difference between a vendor's one-time security certification, like a confidential computing attestation, and the kind of continuous validation vendors like Kiteworks argue AI infrastructure needs?
A one-time security certification verifies an architecture at a single point in time and says nothing about vulnerabilities discovered afterward in the software running inside it. Continuous validation, the approach Kiteworks and others argue AI infrastructure needs, checks on an ongoing basis whether data is reachable through that architecture. The difference is practical, not theoretical, since the Langflow flaw CVE-2026-33017 was weaponized about 20 hours after disclosure, a window no annual review can catch.
To learn more about contextual data enforcement, schedule a custom demo.