The security architecture of modern data centers is built around layers of redundancy, segmentation, monitoring, and access control. Yet beneath the operating systems, hypervisors, containers, and cloud platforms sits a management layer that is often less visible, less monitored, and considerably more powerful than the workloads it controls.
New research has exposed how a vulnerability rooted in a protocol introduced more than two decades ago continues to place thousands of servers—and potentially entire data center environments—at risk.
Security firm Lava identified 36,872 server-management interfaces exposing the Intelligent Platform Management Interface, or IPMI, directly to the public internet. Of those systems, 24,650 disclosed password-derived authentication material before a user had successfully logged in, allowing attackers to perform offline password-cracking attacks against some of the most privileged devices inside a data center. (Lava)
The weakness is tracked as CVE-2013-4786, a high-severity flaw in the IPMI 2.0 authentication protocol. Although the vulnerability was formally documented more than a decade ago, the underlying protocol dates back to 2004, and large numbers of affected management controllers remain reachable from the internet today. (NVD)
The findings demonstrate how a legacy security problem can become more dangerous over time, not because the flaw itself has changed, but because modern computing resources have made exploitation faster, cheaper, and more practical.
At the center of the issue is the Baseboard Management Controller, commonly known as a BMC.
A BMC is a dedicated processor built into a physical server. It operates independently of the main operating system and gives administrators the ability to control the machine remotely, even when the host has crashed, been powered off, or cannot boot normally.
Through a BMC interface, an administrator can restart a server, access its console, mount virtual installation media, modify low-level hardware settings, update firmware, inspect sensors, and perform other operations that would otherwise require physical access to the machine.
Those capabilities make the BMC indispensable for managing large data centers. They also make it one of the most sensitive control points in the entire infrastructure.
Major hardware vendors use different names for their management platforms. Hewlett Packard Enterprise calls its system Integrated Lights-Out, or iLO. Dell uses iDRAC. Lenovo provides XClarity Controller, while Supermicro operates its own BMC environment. OpenBMC, meanwhile, provides an open-source management stack used in parts of the cloud and hyperscale computing ecosystem.
Despite differences between implementations, these systems frequently expose several overlapping management interfaces, including the older IPMI protocol, browser-based administration portals, remote consoles, and newer APIs such as Redfish.
In many deployments, those interfaces share the same credentials.
This means that a password recovered through the IPMI authentication process may also unlock the BMC’s web interface, Redfish API, or other administrative functions. An attacker who begins by exploiting one legacy protocol could therefore gain access to a far broader range of server-management capabilities.
CVE-2013-4786 exists because of the way IPMI 2.0 performs authentication through the Remote Authenticated Key-Exchange Protocol, known as RAKP.
Before authentication has been completed, the BMC can return an HMAC-SHA1 value calculated using the account password and session information already known to the requester. An unauthenticated attacker who can reach UDP port 623 can request this response and take the resulting password-derived material away for offline analysis.
The attacker does not need to repeatedly submit login attempts to the server.
Instead, potential passwords can be tested locally using password-cracking software and modern processors or graphics cards. Because the guessing happens away from the target, it avoids many of the warning signs associated with conventional brute-force attacks, such as large numbers of failed login attempts.
That distinction is crucial.
Online authentication attacks can be slowed through rate limits, account lockouts, network monitoring, and automated alerting. Offline cracking is constrained primarily by the attacker’s available computing power and the complexity of the password.
Lava found that approximately two-thirds of the internet-exposed BMCs it tested returned at least one password-derived RAKP response before authentication was complete. The researchers then compared the captured material against publicly available password lists.
They reported that 6,240 systems accepted an empty username associated with a weak password. Another 2,340 exposed named accounts—such asADMINorroot—whose passwords matched common entries in public wordlists. Many of those passwords were recovered during an initial cracking pass and within minutes. (Lava)
The researchers said they did not use the recovered passwords to authenticate to the affected servers.
Common passwords, however, represented only part of the problem.
Modern server vendors increasingly ship devices with unique passwords rather than using the same default credentials across every unit. At first glance, that practice appears to eliminate the most obvious risk associated with factory-set passwords.
Lava’s research suggests that uniqueness alone is not enough when the password format is predictable and attackers can perform offline cracking.
Supermicro began replacing its traditionalADMIN:ADMINcredential with unique preprogrammed passwords on newer systems in 2019. The company’s documented format consists of exactly ten uppercase letters printed on a label attached to the server chassis.
That produces roughly 141 trillion possible combinations—a search space too large for ordinary wordlist attacks but far smaller than would be created by a longer password using uppercase and lowercase letters, numbers, and symbols.
According to Lava, a modern server equipped with eight GPUs could search the entire Supermicro factory-password space in approximately one hour for a targeted authentication response.
The researchers tested two adjacent internet-facing systems operated by a United States-based bare-metal GPU provider whose published policy allowed limited security testing. Both were modern Supermicro servers manufactured in 2023, and in both cases the researchers recovered passwords matching the vendor’s ten-character factory format.
They did not use the credentials to log in or modify the systems and reported the exposure to the provider, which subsequently addressed the issue. (Lava)
A separate experiment involving Lava’s own HPE iLO server produced an even more dramatic result.
HPE factory passwords used a format consisting of eight uppercase letters and digits, creating approximately 2.8 trillion combinations. On an Apple M3 computer, the researchers said a complete search took roughly one day for a single captured response.
On a system equipped with eight Nvidia RTX 6000 Pro GPUs, the same process took approximately 32 seconds. (Lava)
The comparison illustrates how advances in GPU computing have transformed the practical severity of older cryptographic and authentication weaknesses.
A password policy that may once have appeared sufficient against brute-force attacks can become vulnerable when the authentication protocol allows attackers to capture reusable material and test guesses at extremely high speed.
The issue is especially relevant to artificial intelligence infrastructure.
AI data centers increasingly concentrate large numbers of expensive GPU servers behind shared provisioning platforms, storage systems, high-speed interconnects, and out-of-band management networks. Individual machines may be allocated to separate customers, but the underlying management plane often remains under the control of the infrastructure provider.
A compromised BMC can therefore represent more than control over one server.
It may provide a foothold into shared administrative networks, expose credentials reused across neighboring systems, reveal provisioning infrastructure, or enable movement toward other management interfaces. In poorly segmented environments, the compromise of a single controller could expand into a broader attack affecting multiple servers or even multiple tenants.
This risk is particularly significant for so-called neocloud and specialist GPU providers, which have expanded rapidly to meet demand for AI computing capacity.
Customers may believe they are renting isolated bare-metal machines, while the BMCs, orchestration services, credential stores, and provisioning tools supporting those servers remain connected through shared provider-managed infrastructure.
The operating system running on the leased server may be isolated. The management plane beneath it may not be.
BMC compromise also creates a visibility problem for defenders.
Most enterprise security products monitor activity inside the operating system. Endpoint detection tools inspect processes, memory, files, kernel behavior, network connections, and user activity. Cloud security platforms monitor workloads, containers, identities, and APIs.
The BMC exists outside many of those controls.
An attacker operating through the management processor may be able to access the server console, modify firmware, mount virtual media, or change low-level settings without generating the telemetry normally produced by activity inside the host operating system.
Persistence at this layer can also survive conventional remediation.
Reinstalling the operating system, replacing disks, deleting virtual machines, or rebuilding applications may not remove malicious changes made to BMC or platform firmware. Restoring trust could require verified firmware reflashing, hardware integrity checks, vendor-assisted recovery, or replacement of components that can no longer be considered reliable.
The research also uncovered evidence suggesting that exposed BMC interfaces are already attracting attackers.
Lava found an internet-accessible HPE iLO 4 login page whose security notice had been modified to display a ransom message. The notice claimed the server’s data had been encrypted and demanded 0.3 Bitcoin.
The researchers could not confirm whether encryption had actually occurred, but the altered page indicated that an unauthorized party had gained sufficient access to modify the management interface. (Lava)
Earlier incidents have already shown that this layer can support destructive and persistent attacks.
Security researchers previously documented iLOBleed, a rootkit targeting HPE iLO 4 firmware that could remain hidden from the operating system and survive many conventional recovery procedures. Such cases demonstrate that BMC security is not merely a theoretical concern involving forgotten administrative pages. It is a potential route to control beneath the software stack that organizations typically monitor.
The persistence of CVE-2013-4786 also exposes a difficult reality about infrastructure security: not every dangerous vulnerability can be resolved through a conventional software patch.
The flaw is tied to the design of IPMI 2.0 authentication. While vendors can disable vulnerable functionality, restrict network exposure, or improve default configurations, eliminating the underlying risk often depends on how organizations deploy and manage the protocol.
That places much of the responsibility on data center operators.
Lava’s primary recommendation is straightforward: IPMI should never be directly accessible from the public internet.
Organizations should block UDP port 623 at the network perimeter and restrict BMC access to dedicated private management networks. Remote administration should pass through controlled pathways such as a VPN, bastion host, privileged access gateway, or similarly protected system.
Factory-issued passwords should be replaced during provisioning, even when each server ships with a unique credential. Legacy features such as IPMI 1.5, cipher suite 0, anonymous users, empty accounts, and authentication modes that advertise no password requirement should be disabled.
Network access controls should ensure that only explicitly authorized administrative systems can communicate with BMC interfaces.
Management networks should also be monitored separately from production workloads.
This is an area where many organizations remain exposed. Out-of-band networks are often treated as isolated and therefore inherently trusted. As a result, they may receive less logging, behavioral analysis, asset discovery, or intrusion detection than the application environments they manage.
The assumption becomes dangerous once an attacker reaches that supposedly trusted layer.
Where supported, operators should favor modern management APIs such as Redfish over TLS, but only within a properly isolated administrative environment. Moving from IPMI to Redfish does not justify exposing server management directly to the public internet.
Lava reported its findings to affected operators and hardware vendors.
Supermicro acknowledged that recovery of its factory-password format was plausible using modern multi-GPU systems. The company reiterated that customers should change default BMC passwords during installation and protect management interfaces behind dedicated VLANs and access-control lists. It also said it would evaluate stronger factory-password formats for future hardware revisions.
HPE said it was assessing stronger default-password requirements as well. The company noted that IPMI-over-LAN is disabled by default in newer generations of iLO, reducing the chance of accidental exposure on recently deployed hardware.
A substantial installed base of earlier iLO systems remains in operation, however, and hardening those devices continues to depend on customers following security guidance that has existed for years. (Lava)
That gap is at the center of the problem.
The technical weakness has been publicly understood for more than a decade. Vendors have issued recommendations. Security researchers have described the danger. Tools for identifying vulnerable systems have existed for years.
Yet tens of thousands of interfaces remain publicly reachable.
The persistence of the exposure suggests that advisory-based security does not scale effectively across an infrastructure ecosystem containing countless operators, hosting providers, enterprises, universities, laboratories, and smaller data centers.
Some administrators may not know their BMC interfaces are internet-facing. Others may inherit older systems with unchanged configurations. In large environments, servers can be deployed faster than security teams can verify every management interface. Equipment may remain operational for years after the staff responsible for its original configuration have left.
The result is a layer of critical infrastructure protected by controls that are often optional, manually applied, and inconsistently monitored.
The vulnerability also challenges a common assumption that older security issues naturally become less relevant as technology advances.
In this case, the opposite has occurred.
The protocol has remained in service, while GPU hardware has made password recovery dramatically faster. AI infrastructure has increased the value of the servers behind these interfaces. Greater automation and shared management environments have expanded the potential blast radius of a successful compromise.
A flaw introduced in 2004 may therefore be more operationally dangerous in 2026 than it was when first documented.
The findings are not simply a warning about one protocol or one generation of server hardware. They expose a broader weakness in how organizations secure the management plane beneath modern computing infrastructure.
Data centers often invest heavily in protecting applications, customer workloads, operating systems, cloud identities, and production networks. The systems capable of powering those machines off, rewriting their firmware, or mounting arbitrary media may receive far less attention.
That imbalance gives attackers an opportunity to move below the layers defenders are watching.
The most important lesson from the research is that out-of-band management cannot remain out of sight.
BMCs should be treated as critical security boundaries, protected with the same rigor applied to privileged cloud accounts, domain controllers, hypervisors, and production orchestration systems. Their exposure should be continuously assessed, their credentials rotated, their networks segmented, and their activity monitored for signs of unauthorized access.
The servers powering the cloud and the artificial intelligence boom may run some of the most advanced software ever created. Yet thousands of them remain reachable through a management protocol designed more than twenty years ago.
That contrast captures one of cybersecurity’s most persistent problems: the future is often built on infrastructure whose oldest weaknesses were never fully removed.