A twenty-year-old vulnerability in the IPMI protocol continues to put tens of thousands of servers at risk. Researchers at security firm Lava discovered that more than 24,000 publicly accessible Baseboard Management Controllers (BMCs) expose enough authentication data to allow passwords to be cracked offline. AI infrastructures, in particular, are at increased risk.
The vulnerability, registered as CVE-2013-4786, lies in the authentication mechanism of IPMI 2.0, a protocol used for remote server management since 2004. According to BleepingComputer, an attacker need only request an authentication response from a BMC. That response contains enough information to subsequently crack the password offline using GPUs or other powerful hardware.
Security specialists at Lava scanned publicly accessible IPMI services on UDP port 623 and found 36,872 systems. For 24,650 of these, sufficient authentication data was available to launch an offline attack.
The researchers also found a striking number of poorly secured systems. For example, 6,240 BMCs accepted a blank username during login. Another 2,340 servers were found to be using administrator passwords that appeared in public word lists. According to a live map reviewed by BleepingComputer, the United States tops the list with approximately 39 percent of all vulnerable servers.
A large portion of the exposed systems consists of Supermicro servers. By default, these use the username “ADMIN” and a password consisting of ten uppercase letters printed on a sticker on the chassis. Despite its apparent complexity, this fixed structure makes offline cracking quite feasible.
Risks for AI environments
A BMC operates independently of the operating system and enables low-level management, such as updating firmware or remotely powering servers on and off. Anyone who gains control of the BMC can therefore also manipulate the physical server.
Lava warns that the impact on AI environments could be particularly severe. “In practice, compromised login credentials often work on multiple management interfaces within the same environment. A single compromised BMC can therefore serve as a springboard to the broader management environment,” the researchers say.
They also highlight the implications for shared GPU infrastructure. “A physical GPU server can support multiple tenants or workloads through virtualization, GPU partitioning, or other forms of shared use. In such environments, the compromise of a single physical server can lead to disruption or exposure of multiple customers’ workloads.”
Ransom note found
During the investigation, Lava stumbled upon a publicly accessible HPE iLO 4 interface displaying a ransom note demanding payment of 0.3 bitcoin. According to the researchers, this does not prove that the vulnerability is being exploited on a large scale, but it does show that management systems are actively being targeted by attackers.
After Lava’s report, Supermicro acknowledged the risk and advised administrators on the importance of changing default BMC passwords and isolating management networks. The company is also exploring the use of stricter default passwords for future hardware. HPE also received a report, but according to Lava, it responded only with an automated acknowledgment.
Lava advises organizations never to expose IPMI and Redfish interfaces directly to the internet, to immediately change default passwords, to isolate management networks, and to disable outdated IPMI authentication. A compromised BMC can have far-reaching consequences, particularly in AI infrastructures with shared GPU servers.