DEV Community

Anoymask
Anoymask

Posted on

NoisyTCP CVE-2026-99288: Pre-Authentication Remote DoS Causing Linux Kernel Panic via Crafted TCP Packets

1. Basic Information

2. Executive Summary

NoisyTCP is a denial-of-service (DoS) vulnerability that exploits a race condition between the lockless TCP socket lookup and socket reuse in the Linux kernel. It allows an unauthenticated remote attacker to trigger a NULL pointer dereference and a kernel panic using only crafted TCP traffic. TCP listeners with default configurations, including Nginx, Apache, and SSH, can be targeted, but the reproduction results demonstrated in public materials are limited to a remote DoS causing a host crash. The researcher stated that they created a highly restricted, lab-specific exploit for VMs, but did not demonstrate general-purpose remote code execution (RCE).

3. Attack Flow

Request_sock Reuse ABA Race

  1. The attacker leverages the difference in receive hashes between IPv4 fragments and normal TCP packets to distribute packet processing across different receive (RX) contexts.
  2. One RX context performs an ehash lookup for an existing request socket and holds onto its old address.
  3. Another RX context releases the socket and reassigns a SYN-cookie request to the same tuple and slab address.
  4. The stale lookup acquires a reference to the new object and passes the tuple re-validation, but the new request is unhashed and lacks a listener.
  5. tcp_v4_rcv() references a NULL listener, leading to a kernel panic and halting the entire host.

4. Attacker Position and Execution Location

  • An unauthenticated remote attacker capable of sending packets directly to a TCP listener on the target Linux host.
  • Targets include standard TCP listeners such as Nginx, Apache, SSH, and proxies. The attack occurs within the kernel network stack before any application-level authentication or request processing takes place.

5. Impact on Victims and Administrators

  • Users: Services hosted on the system, including web apps, SSH, and container workloads, suddenly become unavailable.
  • Administrators: Can observe kernel Oops or panics, NULL dereferences in tcp_v4_rcv, high volumes of TCP or fragmented traffic directed at the same listener, and unexpected host reboots. Even if the target is a service within a container, the entire host kernel may crash.

6. Success and Failure Conditions

Success Conditions

  • Attacker-sent TCP packets reach the target directly.
  • Packet processing is handled concurrently across different RX contexts due to multiple RX queues or Receive Packet Steering (RPS).
  • The ehash lookup of the request socket overlaps with its release and reassignment to the same tuple within the race window.
  • For the representative flow that uses the SYN cookie path, SYN cookies must be enabled.

Failure Conditions and Mitigations

  • Apply the vendor-provided patched kernel or an equivalent upstream fix (4e2b3ef19e0e).
  • As a temporary mitigation, limit external interfaces to a single RX queue to serialize processing, and disable both RPS and Receive Flow Steering (RFS).
  • Disabling SYN cookies can also block the typical path, but this reduces SYN flood protection and requires configuration changes per network namespace.

7. What Happens Upon Success

  • Kernel panic and complete loss of host availability.
  • Simultaneous outages of web servers, SSH, proxies, and container workloads running on the host.
  • If the system remains unpatched and externally accessible after a reboot, the same DoS attack can be repeated.
  • Publicly demonstrated reproduction results are limited to a remote DoS (host crash). The researcher mentioned creating a highly restricted, lab-specific exploit for VMs, but did not demonstrate general RCE.

8. Observable Logs

  • Email: None applicable.
  • Proxy / SWG / DNS: HTTP proxy logs may be absent because the attack occurs at the TCP stack level before application handling.
  • Endpoint / EDR: Observe kernel messages such as BUG: kernel NULL pointer dereference, addresses near 0x12, tcp_v4_rcv, Kernel panic - not syncing, and unexpected reboots.
  • Identity / IdP: Because the attack succeeds prior to authentication, successful or failed login events are not required for exploitation, so corresponding authentication logs may be absent.
  • SaaS / Cloud: Cloud VM health check failures, instance reboots, and simultaneous outages of containers on the same host.
  • Network: Traffic combining IPv4 fragments and normal TCP packets, flows distributed across multiple RX queues, and high socket churn. Packet patterns alone do not confirm successful exploitation.

9. Determining Attack Success

Confirmed in Public Information

  • Lab Reproduction Confirmed (Remote DoS Only): The researcher reproduced NULL dereferences and kernel panics from remote traffic in a lab environment. Reports indicate it took about 5 seconds to 2 minutes on a VM running default Nginx, Apache, or SSH configurations, and approximately 14 hours on a quiet bare-metal Nginx host, even with a PoC tuned for that environment. This represents a lab-induced DoS, not confirmation of code execution or actual exploitation in the wild.

Criteria for Internal Assessment

  • Do not assume success based solely on TCP or fragmented traffic; correlate panic stacks, crash timestamps, network captures, host reboots, and service health.
  • Following a crash, avoid assuming other impacts derived from memory corruption; distinguish between the DoS confirmed by public research and additional artifacts specific to your environment.

10. Investigation Playbook

  • Investigation Starting Point: Unexpected kernel panics on externally facing Linux hosts, tcp_v4_rcv stack traces, and simultaneous outages of multiple services.
  • Initial Checks: Verify kernel versions, vendor fixes, SYN cookies, RX queues, RPS/RFS settings, listeners, and immediately preceding traffic.
  • Endpoint / Server Inspection: Preserve dmesg, persistent journals, kdump/vmcore files, boot history, and NIC queue configurations.
  • Authentication / Cloud Inspection: Check cloud instance health, load balancer flows, and reboot events. The absence of authentication logs should not be treated as proof of no attack.
  • Follow-up Actions: Check for repeated attacks from the same source, similar DoS events on other hosts, and availability impacts; distinguish these from lateral movement.
  • Containment: Prioritize patch deployment. If necessary, temporarily apply ACLs or DDoS protection, or limit the external interface to a single RX queue and disable both RPS and RFS.
  • Categorization: Separate incoming traffic, suspected race triggers, confirmed kernel panics, host/service impacts, and recurrences.

11. Defense and Detection Ideas

  • Single Events: Treat tcp_v4_rcv NULL dereferences, kernel panics, and simultaneous health check failures across multiple services with high priority.
  • Timeline Correlation: Correlate fragment/TCP bursts, kernel Oops, panics, instance reboots, and service recoveries on a per-host basis.
  • Threat Hunting: Inventory externally reachable TCP listeners, configurations with two or more RX queues or RPS enabled, unpatched kernels, and recent unexpected reboots.
  • Log Limitations: Logs may fail to flush to disk before a panic occurs. Combine remote logs, network flows, hypervisor/cloud health checks, and kdump.
  • Priority Actions: Prioritize vendor fixes, reducing external exposure, configuring kernel crash dumps, enabling remote logging, and ensuring redundancy. Treat disabling SYN cookies as a temporary measure after evaluating trade-offs.

12. Facts / Inference / Hypothesis

Facts

  • CVE-2026-99288 is a race condition involving lockless socket lookup and socket reuse in the Linux kernel TCP subsystem.
  • When a stale ehash lookup returns an unhashed request socket, a kernel panic occurs due to a NULL rsk_listener reference.
  • The researcher reported that standard TCP listeners, including default configurations of Nginx, Apache, and SSH, can be triggered remotely without authentication.
  • The vulnerability was introduced in commit 079096f103fa on 2015-10-03 and fixed in commit 4e2b3ef19e0e on 2026-11-03.
  • Reproduction times vary significantly by environment, taking roughly 5 seconds to 2 minutes on VMs and about 14 hours on quiet bare-metal systems even with a tuned PoC.
  • Publicly demonstrated reproduction results are limited to remote DoS. The researcher noted the creation of a highly restricted, lab-specific exploit for VMs, but did not demonstrate general RCE.

Inference

  • In container environments that share the host network as well as internet-facing hosts, an attack against a single listener can lead to the simultaneous outage of multiple workloads.
  • Risk management requires correlating patch statuses, RX configurations, panic telemetry, and availability events, rather than relying solely on signatures.

Hypothesis

No additional hypotheses. Lab PoC success is not treated as confirmation of in-the-wild exploitation.

13. MITRE ATT&CK Mapping

ID Technique Confidence Basis
T1499.004 Endpoint Denial of Service: Application or System Exploitation high Causes a Linux kernel panic via crafted TCP traffic, resulting in a loss of host availability.

14. Unknowns and Additional Investigation

  • Presence of in-the-wild exploitation, required packet volumes and race tuning per target, and practical IPv6 trigger conditions.
  • Backport statuses across various distributions and cloud images, and the impact of specialized NIC offloads and virtualization configurations on reproducibility.
  • The researcher explained that a remote use-after-free (UAF) is theoretically possible via a full child socket ABA path. However, this requires orchestrating complex nested race conditions and is difficult to trigger without specialized techniques. Practicality and impacts beyond DoS remain unconfirmed.

15. Impact on SOCs and Organizations

Linux web servers, SSH services, proxies, and container hosts are widely deployed, and application-level patches alone cannot remediate this issue. Inventory externally reachable listeners and kernel packages, and prioritize vendor fixes. Integrate availability monitoring, remote kernel logs, kdump, and load balancer health checks to ensure that host crashes lacking authentication logs can still be investigated as network-layer attacks.

16. Summary by Role

  • SOC Teams: Correlate TCP/fragmented traffic, tcp_v4_rcv panics, instance reboots, and multi-service outages along a timeline to distinguish actual exploitation from routine crashes.
  • Administrators: Apply vendor-patched kernels. When unpatched, evaluate the trade-offs of switching to a single RX queue, disabling RPS/RFS, or modifying SYN cookie settings.
  • Users: This vulnerability allows servers to be crashed prior to authentication. If a service becomes unavailable, avoid continuous retries and report the issue to operations support.

Top comments (0)