pypdf CVE-2026-59935

Exploitation requires that the target application parse an attacker-supplied PDF with pypdf prior to 6.14.2 and reach the vulnerable code path - specifically page content-stream parsing such as page text extraction. The malicious PDF must contain an inline image whose data is not properly terminated and that declares an ASCII85 (A85) or ASCIIHex (AHx) filter; that specific malformed construct IS the trigger condition. Limiting factors: the impact is confined to denial of service (no code execution or data disclosure), and although the CVSS marks UI:N, real-world remote-unauthenticated exploitation depends on the application automatically ingesting untrusted PDFs - interactive desktop use would require a user to open the file. Any deployment that only processes trusted, internally generated PDFs is effectively not exposed.

The CVSS 4.0 vector (AV:N/AC:L/AT:N/PR:N/UI:N, VC:N/VI:N/VA:H) scores 8.7 (High) and cleanly reflects a pure availability impact: no confidentiality or integrity loss, no privileges, no user interaction assumed, low attack complexity. That profile is consistent with the description - a malicious PDF is all that is required, and any server-side pipeline that auto-processes uploaded documents can be triggered remotely without authentication. The real-world caveat is that the network/UI:N rating assumes the target automatically ingests attacker-controlled PDFs; in a purely interactive desktop use case a user would have to open the file, so effective exploitability depends on deployment. EPSS is not provided and there is no KEV listing or known public exploit, so likelihood signals are limited, but the impact is a hard single-threaded hang rather than a transient error, and it is easy to weaponize. Net assessment: a legitimate priority for any service that accepts and parses untrusted PDFs with pypdf, though it is a DoS ceiling only - no code execution or data exposure.

An attacker uploads a crafted PDF to a web service that uses pypdf for text extraction (e.g., a resume parser, invoice ingester, or search indexer). When the backend calls page text extraction on the file, the ASCII85/ASCIIHex inline-image decoder enters an infinite loop, consuming a CPU core and stalling the worker; repeated uploads exhaust the worker pool and take the service offline. No authentication or user interaction is needed against an auto-processing pipeline, and no public exploit is currently required since the malformed structure is simple to construct.