NiBabel CVE-2026-118590

The attacker must get a nibabel-using application to load an attacker-controlled neuroimaging file (e.g. via load()/get_fdata() or a direct array_from_file call) whose header declares a far larger data array than the file actually contains, over an uncompressed stream - on which the pre-patch code seeks no size check and allocates a buffer for the header-claimed shape. Limiting factors: impact is limited to availability (memory exhaustion / process crash), not data exposure or code execution; exploitation requires the victim to parse the malicious file, so true remote reach depends on a network-facing service accepting file uploads rather than the standalone library; and per the PR, compressed streams are still not checked up front but surface a short-read OSError when read.

This is a moderate-priority availability issue, not a critical flaw. The CVSS 4.0 vector (AV:N/AC:L/AT:N/PR:N/UI:N) scores the reachability generously as remote and unauthenticated, but the only impact metric set is VA:L - limited availability, with VC:N and VI:N confirming no confidentiality or integrity consequence. CWE-789 (uncontrolled memory allocation) and the PR diff corroborate that the outcome is memory exhaustion / denial of service, not code execution or data exposure. No EPSS score and no CISA KEV entry are provided, so there is no evidence of active exploitation; the decision label is 'this_month' (not 'act_now' or 'emergency'), consistent with a real-but-not-urgent fix. A public PoC reference exists (GitHub issue #1566 and the reproducer tests in PR #1587); it is not verified as working weaponized exploit code. A key caveat: nibabel is a file-parsing library with no network listener of its own, so the AV:N 'remote' characterization only holds when a network-facing application feeds attacker-controlled neuroimaging files to nibabel - verify against the actual deployment before treating this as remotely exploitable. The 'Information Disclosure' tag contradicts the vector (VC:N) and is incorrect for this availability-only bug.

An attacker crafts a NIfTI file whose header declares an enormous data shape (e.g. 32767×32767×1000 float64) while the file body is truncated, then supplies it to an application that loads neuroimaging files with nibabel. On an uncompressed stream, the pre-patch array_from_file allocates a buffer sized to the header-claimed shape before checking the file actually contains that many bytes, exhausting memory and crashing the process. A public PoC reference and reproducer tests exist in the upstream issue/PR, but no verified weaponized exploit is reported.