# Twenty Years of Exposure: 24,000 Servers Are Handing Out Their Password Hashes to Anyone Who Asks
Somewhere on the internet, an HPE server's management interface is displaying a ransom demand for 0.3 Bitcoin. The attacker didn't need a zero-day. They didn't pivot through a compromised application or exploit a misconfigured cloud bucket. They walked in through a door that's been unlocked since 2004, using a technique publicly documented in 2013.
That's the state of BMC security in 2026.
Researchers at Lava, a cybersecurity startup, scanned for internet-exposed IPMI services on UDP port 623 and found 36,872 accessible hosts. Of those, 24,650 were actively leaking the kind of authentication material that lets an attacker crack passwords offline — at their leisure, on commodity hardware, without ever triggering a login failure alert on the target.
The vulnerability is CVE-2013-4786. The protocol flaw underneath it is older still, baked into IPMI 2.0 when it shipped in 2004. When an attacker initiates an IPMI authentication handshake, the server helpfully responds with a salted hash of its own password before verifying who's asking. You don't need to be authorized to collect it. You just need a packet.
## What a Compromised BMC Actually Means
Baseboard Management Controllers are the quiet layer of server infrastructure that most security teams don't think about until something goes catastrophically wrong. A BMC sits on the motherboard, runs its own processor and firmware, maintains its own network interface, and operates completely independently of whatever OS is installed above it. Power the server off. Power it back on. Mount a virtual USB. Flash new firmware. All from a browser window, anywhere in the world.
This is the feature. It's also the catastrophic attack surface.
When an attacker owns a BMC, they own the physical server at a layer that EDR agents, host firewalls, and SIEM platforms simply cannot see. Malicious firmware persists through OS reinstalls. Configuration changes happen below the hypervisor. And because management credentials frequently reuse patterns across an organization's infrastructure, a single compromised BMC can become a pass into every other server in the management plane.
The Lava researchers noted something that deserves emphasis: in real deployments, recovered BMC credentials often work across multiple management interfaces in the same environment. One server. Whole rack. Potentially the entire data center.
## The Numbers Are Damning Enough Without Spin
Six thousand two hundred forty hosts accepted an empty username during the authentication handshake. Let that sit for a moment. Not a weak password — no username at all.
Another 2,340 instances were protected by passwords matching public wordlists. Dictionary attacks. The most pedestrian form of credential cracking, and it worked on thousands of production servers.
The United States accounts for 39% of the vulnerable population. For Supermicro systems — which represent a significant portion of the exposed hosts — the factory default is a ten-character uppercase password printed on a chassis sticker, with the username ADMIN. The constrained character set makes brute-force practical even without the sticker. The Lava researchers estimated that recovering an HPE factory password takes roughly one day on an Apple M3. A GPU rig cuts that to minutes.
These aren't edge cases. These are production servers, exposed to the public internet, running management interfaces that were never meant to face the open web.
## AI Infrastructure Raised the Stakes
Here's the angle that most coverage is underplaying: a significant share of the servers running the infrastructure underneath the current AI boom are GPU clusters. GPU clusters require physical servers. Physical servers have BMCs. And AI infrastructure, especially in the multi-tenant cloud and colocation arrangements that dominate the market, frequently runs multiple customer workloads on a single physical machine through virtualization and GPU partitioning.
Compromise one physical server's BMC, and you're not just touching one customer's workload. You're potentially disrupting or exposing every tenant sharing that hardware. The blast radius of a BMC compromise in a multi-tenant AI environment isn't one VM or one container — it's the physical substrate that everyone is renting.
The timing is not incidental. The explosion of GPU server deployments over the last three years has dramatically expanded the population of hardware that runs IPMI, often configured by teams whose expertise runs to Kubernetes and PyTorch, not server-room BMC hygiene.
## Vendor Response: Technically Accurate, Practically Insufficient
Supermicro was notified in June. The company acknowledged the risk and pointed to existing guidance: rotate default passwords, isolate management networks. They said they'd review stronger default password policies for future hardware revisions.
That's not a response. That's a liability-limiting statement.
The credential rotation and network isolation advice is correct and has been correct for twenty years. The servers Lava found exposed on the public internet in 2026 prove that this advice is not being followed at scale. Pointing administrators back at documentation that hasn't worked is not a remediation strategy.
HPE received notification and returned a standard acknowledgment. The ransom note on their exposed iLO 4 interface was not, as Lava carefully noted, proof of widespread successful exploitation — but it is proof that someone is actively working this attack surface.
## What Defenders Should Do This Week
The fix is not complicated. The barrier is organizational, not technical.
---
## HackWire Analysis
What strikes me about this story isn't the vulnerability — CVE-2013-4786 is thirteen years old and has been on every serious server hardening checklist for most of that time. What's striking is the shape of who's exposed and why this particular issue is resurging now.
The conventional wisdom in security is that the hard problems are the new ones: novel zero-days, sophisticated supply chain attacks, AI-assisted phishing. And yet here are 24,000 servers, many of them almost certainly running sophisticated workloads, exposed by a protocol design decision made in 2004 that nobody got around to fixing at the configuration layer.
This fits a pattern I've watched repeat across infrastructure security for a decade: the gap between what security teams prioritize and what attackers actually exploit widens every time a new technology wave hits. Right now, that wave is AI infrastructure. The people deploying GPU clusters at scale are optimizing for throughput and uptime. BMC hygiene is not their mental model.
The multi-tenant risk Lava flagged is the part that should worry the cloud-adjacent AI players most. A single compromised BMC in a colocation environment isn't just one customer's problem — it's a potential horizontal movement path through physical infrastructure that was never designed with adversarial tenancy in mind. The cloud security model assumes hypervisor isolation. Below the hypervisor, in firmware and management controllers, that model has edges that nobody is auditing.
The ransom note on the HPE iLO interface is a data point, not a trend — but it signals that this attack surface has graduated from theoretical to operational for at least some threat actors. The question is how long before that changes.
— HackWire Editorial
---
## Related Coverage