Security researchers scanned the public internet for exposed server-management interfaces and found tens of thousands of data center systems, including AI and GPU infrastructure, still vulnerable to a well-documented 2013 authentication flaw. The research by security firm, Lava, is a reminder that some of the oldest, least-glamorous parts of the stack — the out-of-band management chips embedded in nearly every server — remain a soft target hiding in plain sight.
The researchers scanned the Internet for exposed Baseboard Management Controllers (BMCs) — the dedicated processors on server motherboards that let administrators remotely power-cycle, reconfigure, or troubleshoot hardware independent of the operating system. The team identified nearly 37,000 internet-reachable systems running IPMI, the two-decades-old protocol used for this kind of out-of-band management.
Of those, roughly two-thirds returned password-derived authentication data before a login even completed — a known weakness tracked as CVE-2013-4786, embedded in the IPMI 2.0 authentication handshake. That data can be captured and cracked offline, without ever triggering a failed-login alert on the target system. More than 30% of the captured hashes matched passwords recoverable through ordinary wordlists or predictable factory-label formats, and researchers said active exploitation attempts against exposed BMCs are already occurring in the wild.
The scale is notable, but the underlying bug is not new. What the research underscores is how much exposed, poorly segmented management infrastructure persists across the industry more than a decade after the vulnerability was first disclosed.
BMCs are unusually privileged. They can control server power, mount virtual media, flash firmware, and manage hardware sensors — all beneath the operating system, invisible to the endpoint detection and workload monitoring tools most security teams rely on. A compromised BMC can also enable persistence that survives an OS reinstall or even a disk swap, turning incident response into a firmware-level recovery problem.
That risk is magnified in AI and GPU cloud environments, where high-value systems are often clustered behind shared out-of-band management networks. Even customers renting dedicated bare-metal GPU servers may still sit behind a provider-managed management plane that spans multiple tenants. A single compromised controller in that shared layer could, in theory, become a pivot point across customer boundaries — raising the stakes well beyond a single machine.
The core issue isn’t a novel exploit — it’s exposure and configuration drift at scale. The researchers’ recommended mitigations are unglamorous but concrete. Block ports, and replace factory-issued passwords during provisioning rather than leaving them in place. Also, disable legacy or weak authentication options.
For organisations running or renting AI and GPU infrastructure specifically, the research is also a prompt to ask providers pointed questions about how out-of-board management networks are segmented between tenants — since the blast radius of a compromised BMC in a shared environment can extend well beyond the rented hardware itself.
The broader lesson echoes a familiar pattern in enterprise security: the layer beneath the operating system is often the layer receiving the least attention, even as it holds some of the deepest privileges in the data center. A vulnerability first disclosed in 2013 is still yielding results in 2026 largely because secure configuration of BMCs has remained optional rather than enforced by default — a gap that, as this research shows, attackers are already exploiting.