Lava Tells Lets Data Science That Every Compromised Server Controller It Found Was an HPE iLO 4

"Organizations have spent years hardening cloud workloads and operating systems, but many have overlooked the infrastructure that sits beneath them," said Yakir Kadkoda, CTO and Co-Founder at Lava, in the research announcement. "These management controllers hold the keys to servers and data centers. Once compromised, attackers can operate below the visibility of almost any security tools, maintain persistence even after systems are rebuilt, and potentially move deeper into critical infrastructure."

The 36,872 count is a snapshot rather than a running total. Lava told Lets Data Science it was produced through Shodan, "by searching for exposed IPMI services on UDP port 623."

What happened after the snapshot is the more useful finding. "We continued monitoring the exposed population and observed that it changes daily, with endpoints appearing and disappearing as organizations expose or secure management interfaces," Lava said. "On average, we observed approximately 60 newly exposed BMC endpoints per day."

That reframes the problem. This is not a fixed backlog of 36,872 misconfigurations waiting to be cleaned up. It is a population being continuously replenished at roughly 60 a day, which means remediation has to be a standing process rather than a one-off sweep.

Lava's release noted that researchers found ransomware operators leaving extortion notes on exposed interfaces. We asked how many endpoints actually showed compromise rather than mere exposure, and what those notes looked like.

"We observed approximately 50 exposed BMC endpoints that appeared to have been compromised," Lava told Lets Data Science. "All identified cases were HPE iLO 4 endpoints."

The concentration is striking. Not a spread across vendors and generations, but a single product line. Lava attributed the endpoints to two distinct ransomware campaigns, "each displaying a different ransom message through the BMC web interface instead of the normal management console. The messages indicated that attackers had gained administrative control over the BMC and instructed victims to contact them."

An administrator opening what should be a management console found a ransom demand instead.

The most consequential finding in the research is that vendor-assigned credentials no longer hold. Unique factory passwords have long been treated as adequate because they are not shared across devices, but Lava's researchers showed their predictable formats shrink the search space enough for modern GPUs to brute force them. HPE iLO factory passwords were recovered in under a minute. Supermicro's took roughly an hour despite being uniquely assigned per server.

The underlying IPMI protocol weakness is about two decades old. What changed is the hardware available to attack it. And because the protocol allows password-derived material to be collected remotely before login, Lava notes this cannot be solved by choosing stronger passwords.

Supermicro acknowledged the findings and said it will evaluate strengthening future default password policies.

HPE, whose factory passwords fall fastest, did not. "HPE did not provide a direct response to our disclosure," Lava told Lets Data Science.

The company was careful to add context in HPE's favor. During analysis the researchers "observed a significant difference between older and newer iLO generations," finding very few publicly exposed iLO 5 and newer endpoints, "suggesting that newer deployments are much less commonly exposed." HPE also publishes guidance recommending that factory default credentials be changed on deployment. The compromised population Lava found was entirely iLO 4, an older generation.

We asked what share of the exposed controllers belong to cloud and AI infrastructure providers rather than individual enterprises, since that split would show how much of the AI buildout is affected. Lava declined to estimate it.

"We cannot accurately quantify this split because an exposed BMC endpoint alone does not reveal what workloads are running on the underlying server," the company said, whether that is GPUs, AI workloads or ordinary enterprise applications. Researchers attributed some endpoints to cloud hosting and GPU cloud providers through network ownership analysis, but noted that "ASN ownership does not always tell the full picture, as infrastructure may be hosted, resold, colocated, or operated on behalf of another organization."

It is a more useful answer than a confident percentage would have been.

The research is direct about what actually closes the exposure, and it is not credential hygiene alone:

• •Remove BMC management interfaces from public internet exposure entirely
• •Isolate them on dedicated management networks
• •Rotate all factory credentials during deployment
• •Disable legacy IPMI functionality wherever possible

For teams running AI infrastructure, the operational point is the 60-a-day figure. A single audit finds today's exposure. Only continuous monitoring catches tomorrow's.

The full research is available in Lava's BMC report.