Paulos Yibelo reports that a crafted document can escape Ghostscript’s internal sandbox and run commands as its rendering process. Ghostscript interprets PDF and PostScript documents underneath printing, thumbnail and upload workflows. An application can therefore expose the vulnerable renderer without calling it directly or offering a visibly PostScript-specific feature.

CVE-2026-125396 affects upstream Ghostscript before 10.08.0, according to the CVE record. Distribution packages with older upstream numbers can contain backported fixes. The technical article currently carries a November 21 byline, alongside a separate application and disclosure index. The upstream repair shipped earlier, on September 8, so the new discussion is about exposure and patch adoption, not a newly released fix.

Update — 2026-11-22 04:52 EST: source-date correction

The technical article’s current byline says 21 November 2026. Our earlier text assigned November 17 to that article and November 18 to its index. Those publication claims are withdrawn; the source-publication field now follows the technical article’s verified byline. The date this capture was added and its URL are unchanged.

A fresh Debian tracker check confirms the repair versions below, including Bookworm’s 10.0.0~dfsg-11+deb12u9 security package. This correction changes the publication chronology, not the affected-version boundary or the recommended repair.

Evidence status: researcher-controlled demonstrations, with a documented upstream repair. The reviewed sources identify no confirmed campaign exploiting this flaw against victims. pwn.ai publishes the research and sells security services; its application trials remain its own reported results.

A filename can conceal the interpreter that processes it

The flaw involves a use-after-free, access to memory after its object has been released, in Ghostscript’s Type 1 font loader. A crafted document can corrupt the state enforcing -dSAFER, Ghostscript’s built-in access restrictions. That interpreter setting cannot substitute for an independent operating-system boundary after the interpreter itself is compromised.

Yibelo reports different prerequisites for different application paths:

  • Shared CUPS print queue: the tested Debian 13 queue accepted network print jobs without authentication. The demonstration ran a command as the lp service account. Queue reachability and permission to submit jobs matter.
  • Roundcube email preview: viewing a message triggered thumbnail processing in the tested setup. It required the relevant ImageMagick conversion path and vulnerable Ghostscript; the recipient did not click the attachment.
  • WordPress upload processing: Core exposure required media-upload privileges and the Imagick/Ghostscript stack. Separately configured public plugin forms supplied unauthenticated routes in the researcher’s tests.

The index also names applications whose source code merely reaches an image-processing library. A mapped processing path is weaker evidence than a demonstrated complete exploit. It does not establish that every deployment of ComfyUI, Home Assistant or another listed application is compromised or remotely exploitable.

A file extension alone is similarly weak evidence. In the tested decoder chains, content labeled as an image could reach Ghostscript when pixels were loaded or converted.

The renderer inside the container needs the fix

The upstream release, Debian tracker and Ubuntu notice USN-8949-1 provide concrete repair references:

Deployment Verified repair information
Upstream Ghostscript 10.08.0 contains the fix.
Debian 12, Bookworm 10.0.0~dfsg-11+deb12u9 is fixed in the security channel; +deb12u8 remains listed vulnerable.
Debian 13, Trixie First fixed source package 10.05.1~dfsg-1+deb13u2; current tracker lists 10.05.1~dfsg-1+deb13u3 as fixed.
Ubuntu 24.04 LTS USN-8949-1 lists 10.02.1~dfsg1-0ubuntu7.10 for the Ghostscript library packages.
Ubuntu 22.04 LTS USN-8949-1 lists 9.55.0~dfsg1-0ubuntu5.15 for the Ghostscript library packages.

Debian’s current tracker supersedes the research page’s November 12 snapshot, which still describes Bookworm as lacking a fix. Ubuntu’s October 21 notice describes Yibelo’s font-operand finding without a CVE number because it predates assignment. Consult that notice for other Ubuntu branches and support requirements.

For operators, the next checks are practical:

  • Identify the renderer actually used by each worker. Include container images, bundled libraries and document-processing services, not just the host’s installed command.
  • Deploy the applicable vendor-fixed package or rebuilt image. Replace running workers so they use the repaired code. A host update leaves an older container copy unchanged.
  • Limit who can submit work until repaired. Restrict shared print queues and unnecessary upload, import or preview routes. Authentication reduces the submission pool; it does not repair the renderer.
  • Constrain document workers independently. Minimize filesystem access, credentials and network reach. Use supported operating-system isolation alongside application-level checks.

What the thread adds

In the Hacker News discussion, int0x29 asks: “Just say what was found impacted when with versions along with patch dates and versions.” That request points readers toward the current distribution tracker rather than a long disclosure timeline.

megous suggests kernel-enforced restrictions for converters handling untrusted input. That is a defense-in-depth proposal, not evidence that a particular sandbox configuration stops every chain. In reply, pastel8739 proposes WebAssembly without supplying a tested replacement.

From a 34-comment thread, in HN’s own order.

Caveat: the complete application exploit results are researcher-reported; exposure depends on installed renderer packages, conversion policy, input permissions and surrounding system controls.