Skip to content
nuDefend
← Back to blog

SSHole and GPUStack: two vulnerabilities that allowed access without a secret, both requiring an outbound connection for the next stage

·11 min read

Suppose your server runs two common configurations. The first is a musl-based image, such as Alpine, running sshd to provide access to a container or a small server. The second is an inference management platform, such as GPUStack, running behind nginx on the same host. This week, vulnerabilities in both appeared on the same list. According to The Hacker News, on 9 November CISA added SSHole in OpenSSH and the GPUStack authentication bypass vulnerability to its Known Exploited Vulnerabilities (KEV) catalog. The coverage presents them as two separate items on the update list. You need to install both updates, but it is also worth examining the two vulnerabilities together. Both let an attacker in without presenting a secret, and both required the server to open an outbound connection for the next stage of the attack.

What was published

OpenSSH 10.6p2, 5 November. In the release announcement, the OpenSSH team describes a security bypass vulnerability that it classifies as critical. It affects Linux systems that use musl libc and have PAM disabled. According to the announcement, it allows password authentication to be bypassed through sshd when it is built and run without PAM support. The team believes the bug originates outside sshd, but because exploitation passes through it, it can be blocked there. In the same announcement, the team notes that it receives many security reports, many of them generated by or with the help of AI models. It says that some bugs found by AI tools were later discovered independently by another researcher. From this, it concludes that attackers who do not report bugs can probably find them too. For now, the team will therefore publish releases more frequently rather than accumulate fixes until the scheduled release.

SANS ISC, 8 November. Johannes Ullrich reports on data from the SANS SSH honeypots. The honeypots use Cowrie and retain only the first 255 characters of each password. According to the report, the first SSHole exploitation attempt arrived on 5 November, using passwords consisting entirely of the letter A. The number of attempts using long passwords peaked at 713 on 6 November. Ullrich writes that he believes many attempts succeeded in the honeypots and promises to publish further details. According to SANS, the small increase recorded as early as 1 November does not appear to be related to the vulnerability. It came from password-guessing scripts that mistakenly put commands intended to run after login into the password field.

GPUStack, GHSA-gj9m-f8ph-pmhr. According to the advisory, in versions before 2.2.0, the get_current_user function grants the built-in administrator account's privileges to any request identified as coming from loopback, provided force_auth_localhost is set to False. This is the default. There are two ways to reach this state without credentials. In what the advisory describes as the canonical self-hosted configuration, GPUStack runs behind a reverse proxy on the same host. Every remote request therefore reaches it from 127.0.0.1. A forged X-GPUStack-Real-IP header with the value 127.0.0.1 can also cause a request to be treated as local. According to the advisory, the result is full administrative access over the network: modifying users, models and configuration; creating permanent API keys with full privileges that work from any address; and reading worker and cloud provider credentials. Combined with administrator-only capabilities, this leads to code execution on worker hosts. The advisory says the vulnerability was exploited against a version 2.1.1 installation: the attackers created custom backends that ran reverse shells. The severity score is 9.8, and the fix is included in version 2.2.0.

What the coverage adds. The Hacker News attributes to GreyNoise observations of activity following SSHole exploitation: reconnaissance, requests to cloud metadata services, attempts to read password hashes and root's SSH keys, and changes to the authorized_keys file to maintain access. According to the site, in at least one case an attacker identified as GT10 installed TAC Agent, a command-and-control (C2) agent written in Python. It also reports that watchTowr observed the vulnerability being used for botnet and cryptocurrency mining activity. According to The Hacker News, the GPUStack fix was released as early as July 2026.

Access without a secret

For GPUStack, the advisory itself identifies the root cause: authentication relied on the request's network origin. The assumption that “anything from localhost is trusted” sounds reasonable when only local processes can reach the port. A reverse proxy on the same host removes the basis for that assumption: the proxy is a local process forwarding requests from across the internet. The address 127.0.0.1 no longer indicates who sent the request, only which process forwarded it. Relying on the header is even more problematic: any client can set its value.

For SSHole, the release announcement does not detail the mechanism, and there is no reason to guess. But two facts are enough to see the similarity. According to SANS, exploitation attempts used a password consisting of a long sequence of the letter A: input containing no secret. In the configuration without PAM, sshd relied on a response from another component, and that component could be misled.

This is what the two vulnerabilities have in common: the service relied on something else to establish the user's identity—a proxy, a header or an underlying layer—and the attacker could influence it. For every service on your server, it is therefore worth asking not just “Does it require a password?” but also “What does it use to decide whether a request is authorized, and can someone outside influence that decision?” The dangerous answers are familiar: the source address, a header the proxy is supposed to add, a default that trusts local access, an account you assume is locked, and an authentication path that depends on a component you do not check.

An update window alone no longer protects you

According to SANS, the first SSHole exploitation attempt arrived on the same day OpenSSH published the patched release. With GPUStack, the situation is reversed: the documented exploitation targeted version 2.1.1, and according to The Hacker News, the fix had been available since July. Any server that had not been updated remained exposed for months.

One case involves hours, the other months, but the conclusion is the same: there will be periods when a service is vulnerable and the attacker already knows it. The protections that work during those periods are the ones configured before the vulnerability was published.

The second stage connects out from the server

Taken together, the activity documented after initial access forms a list of outbound connections. In GPUStack, according to the advisory, the custom backends ran reverse shells: the worker host opened a connection to the attacker. For SSHole, The Hacker News reports a C2 agent, botnet activity, cryptocurrency miners and requests to cloud metadata services. In each of these cases, the server itself opens the connection.

Not every step an attacker takes requires an outbound connection.

Here you have an advantage. A small SSH server, a container running sshd and a GPU host have few legitimate reasons to connect out: a package mirror, DNS, NTP, a model source and a management server. Everything else can be blocked by default. According to the GPUStack advisory, anyone who gained administrator access could read worker credentials and cloud provider secrets. A worker host that can reach any address on the internet is therefore a good place to start.

Practical steps

OpenSSH

  • Find the systems at risk. On every host, image and container running sshd, check whether sshd is linked to musl: ldd /usr/sbin/sshd 2>&1 | grep -i musl.
  • Check the effective configuration: sudo sshd -T | grep -Ei '^(usepam|passwordauthentication|kbdinteractiveauthentication) '. OpenSSH describes musl-based systems with PAM disabled, and sshd built and run without PAM support. Update to 10.6p2 or your distribution's patched package.
  • Until the update is installed, disable password authentication and allow only key-based authentication. The vulnerability allows password authentication to be bypassed, so disabling it closes this attack path.
  • Do not rely on “locked” accounts. List accounts that do not have a real password hash: sudo awk -F: '$2 !~ /^\$/ {print $1": "$2}' /etc/shadow. Set an explicit AllowUsers list in sshd so that system accounts cannot log in at all, regardless of their status.
  • Look for signs of the documented activity. Check for changes to authorized_keys files, for example with sudo find / -xdev -name authorized_keys -newermt 2026-11-01, and look for unfamiliar processes opening outbound connections. If you suspect compromise, replace root's SSH keys and passwords.

GPUStack

  • Check the version with pip show gpustack. If it is below 2.2.0, update.
  • Until you update, set force_auth_localhost=true so that loopback requests also require authentication.
  • Remove the header at the proxy. In nginx, in the block that forwards requests to GPUStack: proxy_set_header X-GPUStack-Real-IP "";. This prevents a client from passing its own value.
  • Prevent access to the API from untrusted networks, using a VPN or authentication at the edge, as the advisory recommends.
  • If you ran an older version that was accessible from outside, review the users, administrative API keys and custom backends. Replace worker credentials and cloud provider secrets, because the advisory says anyone who gained administrator access could read them.

Any service behind a local proxy

  • Check the configuration of every service behind a reverse proxy on the same host for settings that trust local access or client IP headers. When the proxy runs on the same host, internet requests it forwards to the service arrive as local requests.

Outbound traffic

  • Block by default and allow only what is needed: package mirrors, DNS and NTP on SSH hosts, and the model source and management server on GPU hosts. In systemd, you can use IPAddressDeny=any with IPAddressAllow; in nftables, you can define a rule based on the service's cgroup.
  • Containers that do not need internet access should run on a network with no outbound route.
  • Count blocked outbound connections for each service. An attempt by a worker or SSH host to reach an unfamiliar destination requires immediate investigation.

How nuDefend helps

What the server exposes. nudefend expose shows every listening service on the server, including services inside containers, and how it can be reached: local only, private network, blocked by a local rule, externally accessible or published from a container. An accessible service that is not expected to run on the port where it is listening is flagged for review. When a previously internal service becomes externally accessible, nuDefend alerts you. You can close a port that should not be open from the dashboard or with nudefend block add, after a preview showing exactly what will be blocked. Every block is logged and can be undone with a click.

Which AI tools are running on the server. nudefend workloads shows model servers, chat interfaces, vector databases, automation platforms and MCP servers, on the host or inside a container. It also shows the information used to identify each one. Combined with the exposure map, this lets you see whether any of them are externally accessible.

Where AI tools connect, limited to destinations you approve. nuDefend records which destinations each AI tool on the server connects to, how much traffic was sent to each destination and when the destination was first detected. When a destination is identified, it is shown by name. On the Outbound screen, you approve the destinations the tool actually needs. Once the restriction is enabled, the tool can connect only to destinations you have approved. Every attempt to connect to another destination is blocked before the traffic leaves the server. The rest of the server continues to operate normally. If you do not give final confirmation within ten minutes, the restriction lifts automatically. A restricted tool can also resolve only addresses you have approved. Every other lookup is stopped before it leaves the server and is shown alongside the tool's name.

Threat blocking, in both directions. A cryptographically signed threat list, updated every 30 minutes, blocks traffic to and from known malicious addresses at the kernel firewall. The list includes botnet command-and-control servers, active scanners, sources of password-guessing attempts and malware infrastructure. At the DNS layer, dangerous domains are blocked system-wide, so no application on the server can resolve or reach them. If a process on the server tries to contact a malicious address, the blocked outbound traffic card on the dashboard turns red as an early warning.

Password guessing. Repeated failed login attempts against SSH, mail, FTP and website login pages are blocked at the firewall within minutes. All reports stay on the server.

More details are available on the outbound traffic page, the exposure page, the access blocking page and in the documentation.

Bottom line

SSHole and GPUStack are two different vulnerabilities in two different products, but both failed at the same point. The service did not verify a secret itself. Instead, it relied on something else to decide who could log in: the source address, a header or an underlying layer. Install both updates, and check which other services on your server make that decision in the same way. Because the time between discovery and exploitation ranges from hours to months, configure restrictions in advance that stop the next stage too: limit SSH hosts, containers and GPU hosts to the destinations they actually need.

nuDefend shows what your server exposes and where its AI tools connect, and lets you restrict a tool to destinations you have approved. Connections to any other destination are blocked before the traffic leaves the server.

Check outbound traffic.

Ready to protect your servers with nuDefend?