Linux Kernel CVE-2026-90347

Requires local access and the attacker to be a ptrace tracer of the target process (feasible for one's own children under default YAMA ptrace_scope), AND a seccomp filter that uses SECCOMP_RET_TRACE so that syscall-entry traps re-run the filter after a tracer modification. Exploitation is limited to the arm64 architecture and to deployments that actually rely on seccomp argument filtering or audit for security decisions; it does not affect default configurations lacking a SECCOMP_RET_TRACE tracer, and there is no remote or unauthenticated path (PR:L, AV:L). The consequence is bypass of the first-argument policy check, not direct code execution or memory corruption.

This is a genuine seccomp/audit security-control bypass on arm64 Linux, but the raw metrics overstate it. The vendor CVSS 3.1 base of 8.4 (AV:L/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:N) treats the flaw as if it directly reads and modifies sensitive data. In reality the bug keeps 'orig_x0' out of sync with x0 after a tracer updates x0 during a SECCOMP_RET_TRACE ptrace stop, so seccomp re-evaluation and audit see a stale first argument while the syscall executes with the new one - this defeats argument-based seccomp filtering and audit logging (an integrity/control-bypass problem), not a memory-corruption or information-leak primitive, so the C:H component is questionable and C:L/I:H is more honest. No exploitation-probability signals reinforce urgency: EPSS is absent, this is NOT in CISA KEV (no confirmed active exploitation), and no public PoC is referenced. All exploitation is local (AV:L) and requires the attacker to already hold ptrace capability over the tracee plus a seccomp SECCOMP_RET_TRACE configuration, which sharply narrows realistic impact versus a network-facing 8.4. The 'Information Disclosure' tag conflicts with the actual behavior (a filter/audit bypass) and should be treated with caution. Net: a real defense-in-depth weakness worth patching in sandbox/container hosts, but a lower operational priority than its 8.4 suggests - closer to a mid-7 severity, with no evidence of in-the-wild use.

A local unprivileged user constructs a sandbox that combines a seccomp SECCOMP_RET_TRACE filter with an attacker-controlled ptrace tracer, then, at the syscall-entry stop, rewrites the tracee's x0 register. Because 'orig_x0' is not resynchronized before the seccomp filter is re-evaluated, the argument-based filter and audit subsystem inspect the old value while the kernel executes the syscall with the substituted first argument, bypassing the intended policy. No public exploit code is identified at time of analysis.