Linux Kernel CVE-2026-98382

Requires LOCAL access with both CAP_BPF and CAP_NET_ADMIN. The attacker must load a device-bound-only (dev-bound) BPF/XDP program on a netdev that is offload-registered so the bound-only program inherits a real offdev, then issue bpf(BPF_LINK_CREATE) with a different target ifindex belonging to a sibling port/driver - causing driver-specialized metadata kfuncs to execute against a mismatched driver's xdp_buff (demonstrated as a veth-bound program run on tun). LIMITING FACTORS: this is not remotely exploitable and not reachable from default unprivileged contexts; it depends on delegation of CAP_BPF and CAP_NET_ADMIN (e.g., privileged containers or namespaces) and on the offload/dev-bound BPF configuration being usable. Hosts where only root holds these capabilities, or where XDP offload/dev-bound programs are not used, face substantially reduced exposure.

Signals are mixed and point to a serious-but-constrained local issue rather than an internet-facing emergency. The CVSS 3.1 vector (AV:L/AC:L/PR:L/UI:N, C:H/I:H/A:H, score 7.8) describes a local, low-complexity attack by a caller who already holds privileges, yielding full confidentiality/integrity/availability impact on the host - consistent with the description's out-of-bounds read of a foreign xdp_buff that crashes the kernel (KASAN null-ptr-deref / general protection fault). Exploitation probability is low: EPSS 0.16% (4th percentile), and the vulnerability is NOT on CISA KEV, so there is no confirmed active exploitation. No public exploit identified at time of analysis - the only linked artifacts are the stable-tree fix commits and a VulDB/NVD advisory, not working exploit code. The vuln.today decision label is 'this_week', which aligns with a privileged-local memory-safety bug that merits prompt patching but is not flagged act_now/emergency. Note the practical gating factor the raw score understates: the description states the attacker needs both CAP_BPF and CAP_NET_ADMIN and a specific BPF offload/dev-bound configuration - PR:L captures 'some privilege' but these are substantial capabilities, so the realistic attacker population is narrow (containers/namespaces granting net admin, or multi-tenant hosts delegating BPF). Treat as a priority patch for hosts that delegate those capabilities; lower urgency where only root holds them.

A local actor holding CAP_BPF and CAP_NET_ADMIN (for example within a container or namespace granted network administration) loads a device-bound-only BPF/XDP program associated with one driver, then calls bpf(BPF_LINK_CREATE) targeting a different netdev's ifindex so the program's driver-specialized metadata kfuncs run against an unrelated driver's xdp_buff. Running a veth-bound program against tun causes the kernel to read tun's bare stack xdp_buff as a veth_xdp_buff, producing an out-of-bounds read that oopses and panics the host. No public exploit is identified at time of analysis; the attack requires the specific capabilities and offload/dev-bound setup described.