More than 24,000 internet-accessible servers are actively leaking password hashes because of a 20-year-old flaw built directly into their out-of-band management controllers. This isn't a fresh zero-day or a complex software bug. It's a structural flaw in the Baseboard Management Controller (BMC) specification that has existed since 2004. As organizations rush to deploy specialized hardware for artificial intelligence ai cybersecurity workloads in 2026, these unmanaged legacy endpoints create massive exposure points that undermine host-level defenses.
Security researchers at Lava recently scanned the internet for publicly accessible Intelligent Platform Management Interface (IPMI) services running on UDP port 623. They identified 36,872 exposed hosts. Of those, 24,650 systems sent back password-derived authentication material to unauthenticated callers. That means anyone with an internet connection can ping these hardware controllers and receive password hashes ready for offline cracking.
How Legacy Protocols Fuel AI Cybersecurity Threats
BMCs operate as dedicated microcontrollers embedded on server motherboards. They grant system administrators full out-of-band management authority, allowing them to cycle power, flash firmware, or mount virtual media remotely, regardless of whether the primary operating system is running.
When an out-of-band controller sits directly on the public internet, it bypasses traditional perimeter controls. Attackers don't need to break through host-level endpoint protection or hypervisor guards. Instead, they interact straight with the bare-metal management engine. In modern environments where high-performance compute clusters run critical workloads, exposing these interfaces introduces severe risks. Threat actors who capture BMC access can manipulate hardware configurations, execute silent firmware-level modifications, or cut off remote host access entirely.
The problem escalates in multi-tenant environments. A single physical server often hosts partitioned GPUs or virtualized containers running autonomous AI agent security tasks. If an attacker gains BMC control over a single bare-metal host, they can tamper with low-level execution environments, disrupt co-located client instances, or access raw memory streams across multiple workloads.
The Technical Mechanism Behind CVE-2013-4786
The vulnerability at the center of this mass exposure is cataloged as CVE-2013-4786. It stems from the IPMI 2.0 RMCP+ Authenticated Key-Exchange Protocol (RAKP) specification. During the initial authentication handshake, an unauthenticated client sends an RAKP Message 1 request to UDP port 623.
The protocol specification obligates the BMC to respond with RAKP Message 2. Crucially, this response contains a salted SHA1 or MD5 HMAC password hash for whatever username was requested—before the caller ever proves they possess the valid password.
Because the protocol obligates the server to return this hash prior to authentication, an attacker doesn't need to guess credentials online or trigger account lockouts. They simply request the HMAC hash over UDP, store it locally, and run offline brute-force or dictionary attacks using GPU hardware rigs, Hashcat, or John the Ripper.
Because this behavior is mandated by the core IPMI 2.0 specification, hardware vendors cannot patch the underlying flaw without breaking compliance with the IPMI standard. Fixing it requires disabling legacy IPMI authentication entirely or placing the interface behind strict network isolation boundaries.
Mass Exposure: Weak Passwords and Default Credentials
The research conducted by Lava demonstrates how easily threat actors can turn this architectural flaw into full system compromise. Out of the 24,650 vulnerable hosts leaking authentication hashes, 6,240 instances accepted empty usernames during authentication while remaining protected by weak passwords. Another 2,340 hosts used default administrative credentials that matched public dictionary lists, highlighting the ongoing risk of non-human identity and credential sprawl.
A substantial portion of the exposed BMCs belong to Supermicro systems. These machines ship with a 10-character uppercase password printed directly on the physical chassis label, paired with a static 'ADMIN' username. While a 10-character uppercase string offers theoretical entropy, its predictable structure makes offline GPU cracking remarkably fast. For comparison, researchers estimated that recovering a default HPE factory password from a captured IPMI response requires roughly one day of processing time on an Apple M3 system.
The exposure isn't just theoretical. During their investigation, Lava researchers discovered an internet-facing HPE iLO 4 login interface that had already been defaced with a ransom note demanding 0.3 Bitcoin.
Geographically, the exposure is heavily concentrated. Geographic telemetry revealed that 39% of all vulnerable, internet-exposed BMCs are hosted on networks within the United States.
When notified of the findings in June, Supermicro acknowledged the inherent protocol risks, pointing to existing vendor documentation that instructs administrators to isolate management networks and change default credentials. Supermicro added that it is evaluating stronger default credential mechanisms for future server generations. Meanwhile, HPE returned only an automated response message, leaving security teams without updated guidance for older out-of-band management hardware.
Why Bare-Metal BMC Exposure Threatens AI Agent Security
As enterprise environments transition toward agentic automation and autonomous AI workflows in 2026, physical infrastructure security becomes paramount. Modern AI agent architectures, including agentic defense strategies, rely on strict isolation boundaries between model execution planes and underlying infrastructure.
When attackers gain out-of-band access through a compromised BMC, those isolation boundaries dissolve. An adversary with physical control over the bare-metal layer can execute out-of-band actions that bypass every software security defense configured on the main operating system:
- Bypassing Hypervisor Security: Attackers can access physical hardware control planes directly, making host-level security tools, EDR agents, and hypervisor logging blind to the intrusion.
- Tampering with Agent Workloads: Unauthorized access to BMC firmware allows attackers to intercept hardware data buses, alter persistent storage, or inject malicious payloads prior to system boot.
- Targeting Enterprise Hardware: Legacy hardware deployments from enterprise vendors—including legacy configurations seen across Supermicro, HPE, and IBM enterprise servers—remain vulnerable if out-of-band interfaces were exposed during rapid cloud expansions.
Securing these low-level interfaces requires a complete reassessment of management traffic. Enterprise security architectures must treat out-of-band management networks with the same rigor applied to production core networks.
Actionable Defenses and Cybersecurity Best Practices
Mitigating IPMI exposure demands immediate structural changes rather than waiting for vendor firmware patches. Organizations should align their hardening efforts with standard CISA guidelines for out-of-band device security:
- Enforce Complete Network Isolation: Disconnect all IPMI, HPE iLO, Dell iDRAC, and Redfish interfaces from public-facing IP addresses. Place management interfaces inside a dedicated, isolated out-of-band VLAN accessible only through authenticated bastion hosts or encrypted VPN tunnels.
- Rotate Factory Default Credentials: Immediately update default passwords on all BMCs. Never rely on passwords printed on physical chassis labels or vendor documentation defaults.
- Disable Legacy IPMI Protocols: Where supported, turn off legacy IPMI 2.0 RAKP services completely. Transition hardware management to modern, secure interfaces like Redfish over HTTPS with multi-factor authentication.
- Conduct Complete Asset Auditing: Run internal scans across all IP subnets on UDP port 623 to identify forgotten or misconfigured BMCs. Use audit scripts to verify that no management interface leaks RAKP hashes to unauthenticated subnets.
- Implement Continuous Monitoring: Log all access attempts to out-of-band networks and alert on abnormal traffic patterns directed at management controllers.
By securing the complete physical and virtual stack, enterprise security teams can prevent 20-year-old protocol flaws from compromising modern AI applications and critical enterprise data. For step-by-step guidance on network segmentation and zero-trust administrative access, explore our technical tutorial library covering infrastructure defense strategies.
