Skip to content
nuDefend
← Back to blog

The attack agent needs outbound connections: what three reports on AI attacks tell Linux server administrators

·13 min read

Suppose your server runs an administration interface, a hosting control panel or a model server accessible from the internet. The next attempt to break in is likely to come from something other than a person typing commands. Between October 19 and 21, Wiz, SOCRadar and Symantec published three reports on attacks in which AI agents played a central role, with varying degrees of autonomy. According to Wiz, the agent ran the operation. In SOCRadar's report, the model planned and the operator approved each step. Symantec captured an attack agent's instructions in a honeypot. The environments in the three reports differ considerably: a multilayer cloud environment, ransomware victims' networks and open model servers. But from the perspective of a single Linux server, the pattern repeats: entry is through an internet-accessible service, and many of the channels described later in the attack are outbound connections from the server.

Three reports in three days

Wiz, October 19. The Wiz incident response team describes an investigation in which, it says, an AI agent-based system ran the entire operation from start to finish. One confirmed entry point was a Java remote code execution vulnerability in an exposed Kafka UI instance. Shell commands ran from the application's Java process. Malware was then deployed and a command-and-control (C2) channel established. From there, the agent moved into the cloud, Kubernetes, CI/CD systems and the identity provider. Wiz describes hundreds to thousands of actions per hour, without pauses, and failed scripts repaired within minutes. According to the report, the same source address was observed scanning more than 600 other Wiz customer environments during the same period.

SOCRadar, October 20. The SOCRadar research team found an exposed command-and-control server belonging to affiliates working with the INC and The Gentlemen ransomware groups. Through that server, the team reached six more servers. The team analyzed 8,257 messages across 133 OpenCode conversations using commercial models through OpenRouter. According to the analysis, the AI carries out the work and the human approves it: the model plans, generates commands and summarizes data, while the operator approves each step. Entry into victims' systems was through endpoints and remote access services, including cPanel/WHM. The operators' records describe more than 49 compromises of small hosting providers, but SOCRadar stresses that these records were not independently verified. C2 and tunneling tools were found on the servers, including Ghost C2 and RevSocks, along with 46 versions of the INC encryptor for Windows and Linux. Some coverage, such as LavX News, presents exploitation as a verified fact and notes that code was developed to exploit a LiteLLM vulnerability. According to SOCRadar, that code was only developed or mentioned; its use was not proven.

Symantec, October 21. Symantec researchers captured in a honeypot the complete instruction set of an attack agent. The agent searches the internet for open language model endpoints and uses them as its reasoning engine, at their owners' expense. The instructions allowed it to take up to 100,000 steps without human approval and included about 74 tool definitions for scanning, exploitation and credential theft. They also instructed it to install cryptocurrency miners covertly. Another request arrived the next day from a different address. It contained a script that installs XMRig and configures switching between four mining pools. According to the researchers, the honeypot recorded more than 127,000 external requests from tens of thousands of distinct addresses during the monitoring period.

The entry point: a service that should not have been open

All three reports describe services left accessible from the internet. In Wiz's report, this was Kafka UI; the API of an internal development tool was also probed and tested. SOCRadar lists cPanel, SimpleHelp and FortiClient EMS. Symantec lists model servers, AI gateways and MCP servers. The tool catalog of the agent described by Symantec also contained exploit modules for VS Code, Jupyter, Django, Laravel, Supabase, Next.js and Flask/Werkzeug, along with a dedicated tool for extracting data from databases exposed through Prisma Studio.

According to Symantec, most AI tools now use the same OpenAI-compatible interface. A single scanner can therefore find many types of targets using the same sequence of checks. The researchers recommend assuming that an exposed AI service will be found and probed shortly after it becomes accessible. Wiz adds an important detail: even after gaining a foothold, the agent continued scanning and looking for additional entry points in parallel. Every exposed service is another entry point the agent will try, whether it is the first or the fifth.

Once inside, almost everything connects out

Taken together, the channels described in the three reports form a list of outbound connections. Wiz describes a C2 channel, results sent to the request-capture service webhook[.]site, data transferred to the organization's own code repository and attempts to create storage buckets. The bucket creation attempts were denied. SOCRadar describes Ghost C2, which communicates over HTTPS through Cloudflare, Telegram or DNS, and RevSocks, which creates a reverse SOCKS5 tunnel for traffic within the victim's network and to prepare for data exfiltration. It also lists dedicated exfiltration tools, including uploads to object storage using restic. Symantec describes mining software downloads and connections to mining pools.

When the server cannot open a connection to the attacker's chosen destination, most of these channels are blocked. The agent can still read files and environment variables. But the channels described in the reports for exfiltrating data, downloading tools and receiving instructions are outbound connections from the server. The list of allowed destinations therefore needs to be precise. Some channels rely on legitimate services: the organization's code repository, Cloudflare, Telegram and object storage. Blanket permission for Cloudflare or all of GitHub leaves these channels open. A useful list contains only exact hostnames.

The reasoning engine is not running on your server, unless you left it open

In SOCRadar's report, the models were accessed through OpenRouter from the operators' infrastructure, using paid OpenRouter accounts. In Symantec's report, the reasoning engine is someone else's open endpoint. Symantec describes two types of victims. The owner of the exposed model pays for every step the agent takes. The servers the agent compromises have data and credentials stolen and are used for mining. Your server therefore has two possible roles, and it is worth distinguishing between them. To protect it in advance as an attack target, the main measure is to restrict outbound connections. To protect it from being used as a reasoning engine, restrict inbound access: require authentication on every route and configure the model server so it does not listen on all interfaces.

Block one channel, and the agent moves to the next

According to Wiz, data collection and exfiltration continued throughout the attack, adapting to the channels allowed in each environment. When one route failed or reached a size limit, the agent moved to the next. In Symantec's report, the script configured four mining pools so that blocking one would not stop the others. In SOCRadar's report, Ghost C2 has three communication channels. The practical conclusion is that a blocklist of known malicious destinations is useful, but an agent trying one destination after another will eventually find one that is not listed. The more resilient approach works in reverse: a list of approved destinations, with everything else blocked.

Speed determines which defenses work

Hundreds to thousands of actions per hour, scripts repaired within minutes and up to 100,000 steps without approval. At that pace, responding to an alert comes too late. By the time you read it, the agent has already read environment variables, contacted the cloud metadata service and tried the credentials it found on the next system. For an incident at this scale and speed, Wiz recommends rotating all credentials in compromised resources, without trying to determine which ones were actually read. The important defenses here are those configured in advance: fewer exposed services, outbound connections blocked by default and secrets kept out of environment variables.

The noise works in your favor

Wiz describes these attacks as noisy. An agent trying channel after channel quickly produces a sequence of blocked outbound connections from the same process to changing destinations. A human administrator almost never produces this pattern, making it a clear signal for an alert. But the alert is useful only if blocking is already in place and blocked connections are counted.

There is also a second signal, common to all three reports. Each describes activity being presented as authorized work to get a commercial model to cooperate, although in Wiz's case this is only an assessment. Wiz assesses that resources, files and payloads were given names such as pentest, probe, recon, redteam and exfil, probably to bypass the model's safeguards. Repeated attempts used the same numeric identifier with an added letter, such as -b, -c and -d. In SOCRadar's report, when models refused, the operator rephrased requests using "Authorized Red Team Assessment Request" templates. In Symantec's report, the jailbreak presented reading .env files and API keys as routine development work. The disguise that helps the agent bypass model safeguards reveals it to the server administrator. In Wiz's report, it left traces in resource names. A file or resource named pentest on a server where no penetration test is planned warrants immediate investigation.

Practical steps

  • Map what is accessible from outside. Look especially for administration interfaces, development tools, hosting control panels, remote access tools and model servers. Anything that does not need to be public should listen only on localhost or a private network. For AI services, Symantec recommends requiring authentication on every route, including model-listing and health-check routes, and placing the service behind a gateway rather than configuring it to listen on all interfaces.
  • Update what remains exposed. If you run cPanel/WHM, check that the version fixing CVE-2026-41940 is installed.
  • Block outbound connections by default, separately for each service. On Linux, options include IPAddressDeny=any with IPAddressAllow in systemd, or an nftables rule based on the service's cgroup. Approve exact hostnames. Do not include request-capture services or paste sites in the allowed destinations, and do not allow all of GitHub or all of Cloudflare. SOCRadar also recommends restricting tools such as restic and rclone to authorized computers.
  • Restrict DNS queries. SOCRadar's report documents C2 communication over DNS as well. A restricted service should resolve only the names you have approved.
  • Set alerts for blocked outbound connections. Count blocked outbound connections per process. A sequence of blocks to changing destinations within a short period requires immediate investigation.
  • Look for visible signs. Look for file, resource and user names containing pentest, probe, recon, redteam or exfil, and repeated numeric identifiers with an added letter. Also look for shell processes whose parent is a Java service or web service. Symantec describes hidden files under /tmp, configuration files with the immutable attribute and access logs that are repeatedly deleted. The indicator list in Symantec's report includes the mining pool addresses.
  • Move secrets out of environment variables. According to Wiz, the agent searched environment variables for strings such as token, secret, key and pass. Move secrets to a secrets manager and reduce the permissions they grant. On cloud servers, Symantec recommends enabling version 2 of the metadata service (IMDSv2) with a hop limit of one.
  • Monitor AI key usage. If you run a gateway or hold keys for model providers, unusual inference volume or spending is often the first sign that access has been stolen, according to Symantec.
  • If you are compromised, rotate everything. Rotate every credential that was accessible from the server, and check for unusual use after the compromise.

How nuDefend helps

What the server exposes. nudefend expose shows every listening service on the server, including services running in containers, and classifies its accessibility: local only, private network, blocked by a local rule, externally accessible or published from a container. An accessible service running on an unexpected port is flagged for review so it stands out in the service list. When a previously internal service becomes externally accessible, nuDefend alerts you. A port that should not be open can be closed from the dashboard or with nudefend block add, after a preview shows exactly what will be blocked. Every block is recorded and can be undone.

Which AI tools run 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 how each tool was identified. Some of these service types, such as model servers and MCP servers, also appear among the agent's targets in Symantec's report. Combined with the exposure map, this lets you see whether any of them are accessible from outside.

Where AI tools connect, restricted to destinations you approve. nuDefend records the destinations each AI tool connects to, how much traffic was sent to each destination and when it was first detected. When a destination is identified, it is shown by name. At the end of the monitoring period, you approve the destinations the tool actually needs on the Outbound screen. Once the restriction is enabled, the tool can connect only to destinations you approved. Every attempt to connect elsewhere is blocked before traffic leaves the server. The screen shows whether each tool is restricted. If you do not give final confirmation within ten minutes, the restriction lifts automatically.

What the tool resolves, too. A restricted tool can resolve only the names you approved. Every other lookup is stopped before it leaves the server and appears on the screen alongside the tool's name. This is usually the first sign that a tool is trying to reach a destination you have not approved.

The threat layer. A signed threat list, updated every 30 minutes, blocks traffic to and from known malicious addresses in 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 neutralized system-wide. If something on the server tries to contact a malicious destination, the blocked outbound traffic card on the dashboard turns red as an early warning. Repeated password-guessing attempts against SSH, mail, FTP and website login pages are blocked at the firewall. Reports stay on the server.

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

Bottom line

Three reports published in three days describe different agents, attackers and environments, but the same attack pattern. The agent enters through a service that should not have been open, and needs destinations outside the server to make use of that access. When one channel is blocked, it moves to the next, faster than any human response. What matters is what was configured beforehand: few exposed services, outbound connections restricted to approved destinations and monitoring of blocked connections. A noisy agent encountering blocks leaves a series of blocked attempts behind it, and sometimes files named pentest. That is an opportunity to detect it, provided blocking is already in place.

nuDefend shows what your server exposes and where its AI tools connect, and lets you restrict a tool to destinations you approve. Every other destination is blocked before traffic leaves the server.

Check outbound traffic.

Ready to protect your servers with nuDefend?