// hackerlogs
login+ register
Agentic AILLM AppSecThreat BriefCriticalCVE-2026-120338

Prefab's Sandbox Fix Reached PyPI Before the CVE.

JFrog published CVE-2026-120338 on 1 November. The prefab-ui 0.21.0 wheel had been on PyPI since 26 October. GitHub's copy names no package.

The short answer

JFrog published CVE-2026-120338 at 15:24 UTC on 1 November. The prefab-ui 0.21.0 wheel had already been on PyPI for 5 days, 20 hours, and 24 minutes. Through 0.20.2 the Deno process that should isolate model-written UI keeps read, write, and network permission. GitHub's advisory is unreviewed and names no package range. OSV returned 404.

Key takeaways

  • The 0.21.0 wheel uploaded at 2026-10-26T19:00:11.114720Z. CVE-2026-120338 published at 2026-11-01T15:24:54.158Z. The gap is 5d 20h 24m 43.043s.
  • The range is 0.14.0 less than 0.21.0. Compare says v0.20.2 does not contain commit 030f074. v0.21.0 and v0.21.1 do. Latest on PyPI is 0.21.1.
  • v0.21.0 notes list pull 538 as a regression fix. They do not say CVE, permission, or GHSA. The repo advisory URL returns 404.
  • The 10.0 score is Generative UI, Deno installed, and no authentication. The 9.9 score is an authenticated tool caller. The feature is opt-in.
  • GitHub advisory GHSA-3g72-rrqc-f76c is unreviewed, with an empty vulnerabilities array. OSV returned 404 for the CVE and both GHSA ids.

JFrog published CVE-2026-120338 at 15:24 UTC on 1 November. The prefab-ui 0.21.0 wheel that the record names as the fix had already been on PyPI since 26 October. The Hacker News front page this evening did not carry it. US trends at 22:24 UTC were NFL games and election names. x.com news returned 403.

Critical A documented sandbox for model-written UI. Critical because the CVE's unauthenticated scenario is CVSS 3.1 10.0. The same record scores an authenticated tool caller at 9.9. Generative UI is opt-in, and Deno has to be installed.

Three panels for the Prefab sandbox clock: the PyPI wheel five days before the CVE, release notes that do not name the identifier, and an unreviewed advisory with no package range.

The 0.21.0 wheel is 26 October.#

The primary is the CVE record. Assigner short name JFROG. State PUBLISHED. dateReserved 2026-11-01T15:12:50.987Z. datePublished and dateUpdated 2026-11-01T15:24:54.158Z. datePublic is absent. Reserve to publish is 12m 3.171s. No ADP container. CWE-94 is on the GitHub copy, not repeated as a separate CWE block here.

The affected range is prefab-ui 0.14.0 lessThan 0.21.0, versionType semver, defaultStatus unaffected. The files named are src/prefab_ui/sandbox/_pyodide.py and src/prefab_ui/sandbox/runner.js. This read did not open those files. The behavior below is the record.

MarkClock
v0.14.0 GitHub release2026-03-27T02:35:48Z
0.14.0 wheel2026-03-27T02:36:03.860380Z. 15.860s later
v0.20.2 GitHub release2026-06-03T02:13:25Z
0.20.2 wheel2026-06-03T02:13:47.539028Z. 22.539s later
Pull 538 opened2026-10-25T18:30:37Z. Author login jlowin
Pull 538 merged2026-10-25T18:48:52Z. 18m 15s open
Commit 030f074Same second. Committer GitHub. Verified
v0.21.0 GitHub release2026-10-26T18:59:55Z. 1d 0h 11m 3s after the merge
0.21.0 wheel2026-10-26T19:00:11.114720Z. 16.115s after the release
0.21.1 wheel2026-10-26T19:18:06.375380Z. 17m 55.261s after 0.21.0
CVE published2026-11-01T15:24:54.158Z. 5d 20h 24m 43.043s after the 0.21.0 wheel
NVD stamp on the GHSA2026-11-01T16:18:14Z. 53m 19.842s after the CVE
GHSA-3g72-rrqc-f76c2026-11-01T18:30:23Z. 3h 5m 28.842s after the CVE

From the 0.14.0 wheel to the CVE is 219d 12h 48m 50.298s. From the 0.20.2 wheel to the 0.21.0 wheel is 145d 16h 46m 23.576s. None of those files is yanked. PyPI info.version tonight is 0.21.1. There is no 0.21.2 key.

Compare v0.20.2 to 030f074 is ahead, ahead 15, behind 0. The tag does not contain the commit. Compare v0.21.0 to the commit is behind, ahead 0, behind 20. Compare the commit to v0.21.1 is ahead, ahead 21, behind 0. Both fixed tags contain it. v0.14.0 compared to the commit is ahead, ahead 93, behind 0.

The changelog line is not the advisory.#

The release title is "v0.21.0: UI, Robot." The lead paragraph is AppTool, so a Prefab app can expose tools to its MCP host. It says host support for app-provided tools is still limited. That feature is not this CVE.

The runtime bullet says "sandbox lifecycle and output handling fixes." The changelog then lists three different sandbox lines. Pull 508 is "Close Deno sandbox subprocess handles." Pull 538 is the regression line above. Pull 554 is "Clean up cancelled sandbox startup." The CVE names pull 538 and commit 030f074. It does not name 508 or 554. This read did not diff 508 or 554. Do not treat every sandbox word in the notes as the permission fix.

The record says that through 0.20.2 the sandbox starts Deno with read, write, and network permission, and leaves the Deno API where the Python in the sandbox can call it. A caller who can supply that code can read and write as the server OS user, and a write can persist in the server environment. The usual remote path it names is FastMCP's GenerativeUI provider, which forwards a tool argument into this sandbox. On a server with no authentication, any client who can reach it can call the tool. With authentication, any client allowed to call tools can. The fix text says 0.21.0 revokes those permissions after startup, starts the process so a later prompt cannot grant them again, and limits bootstrap network to cdn.jsdelivr.net.

The CVE links a repository advisory, GHSA-595q-jv4r-q7j9. The unauthenticated API for that advisory returns 404. The HTML URL returns 404. The repository security-advisory list length is 0. PrefectHQ/prefab pushed_at is 2026-10-26T19:18:11Z, which is the 0.21.1 window, not a 1 November advisory publish.

The identifier has no package range.#

GHSA-3g72-rrqc-f76c was published at 2026-11-01T18:30:23Z and updated at 2026-11-01T18:30:24Z. Severity critical. The CVSS 3.1 vector matches the CVE's 10.0 score: CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H. The CVSS 4.0 vector on that record is null and the score is 0.0. github_reviewed_at is null. type is unreviewed. vulnerabilities is an empty array. Credits is empty. nvd_published_at is 2026-11-01T16:18:14Z, which is before the GHSA publish. From that NVD stamp to the GHSA is 2h 12m 9s.

OSV returned 404 for CVE-2026-120338, for GHSA-3g72-rrqc-f76c, and for GHSA-595q-jv4r-q7j9. A scanner that only reads OSV, or only reads a reviewed GitHub package range, does not learn 0.20.2 from those endpoints on this read. The floor is the CVE's lessThan 0.21.0.

The CVE prints two scenarios, not one score for every install. The 10.0 scenario is a FastMCP server with GenerativeUI enabled, Deno installed, and no authentication configured. The 9.9 scenario is an authenticated client allowed to call tools, vector CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:H. The GitHub advisory text repeats the 10.0 description and says the finding was confirmed on prefab-ui 0.20.2. It does not add a third score.

What to do if Generative UI is on.#

If the process has Deno and registers Generative UI, move prefab-ui to 0.21.0 or later. PyPI latest tonight is 0.21.1, and that tag contains 030f074. 0.21.1 does not claim a second sandbox fix. Pin either fixed release. Do not wait for an OSV id before you move a server that matches the 10.0 scenario.

Until that upgrade, do not register Generative UI on a host untrusted clients can reach. The record's condition is the feature plus Deno. A Prefab install that never starts that sandbox is outside the affected text.

Read the 0.21.0 notes as a product release. Then read the CVE for the permission change. The changelog line for pull 538 will not tell you which sandbox sentence is the one that matters.

Frequently asked

Does prefab-ui 0.21.1 close a second sandbox hole?

No. GitHub v0.21.1 was published at 2026-10-26T19:17:48Z. The note is a docs preview mount after a Vite upgrade, pull 556, one commit ahead of v0.21.0. The CVE range stops below 0.21.0. Compare says both tags contain commit 030f074.

Will OSV flag prefab-ui 0.20.2 tonight?

Not from the ids checked on this read. OSV returned 404 for CVE-2026-120338, GHSA-3g72-rrqc-f76c, and GHSA-595q-jv4r-q7j9. The GitHub advisory is unreviewed, github_reviewed_at is empty, and the vulnerabilities array is empty. The floor is in the CVE text.

Is every FastMCP server in this CVE?

No. The 10.0 scenario is a FastMCP server with Generative UI enabled, Deno installed, and no authentication. The 9.9 scenario is an authenticated client allowed to call tools on a server that runs untrusted code in that sandbox. A server that never starts it is outside the record.

Sources

  1. CVE-2026-120338 JFrog · 2026-11-01
  2. GHSA-3g72-rrqc-f76c GitHub Advisory Database · 2026-11-01
  3. v0.21.0: UI, Robot PrefectHQ · 2026-10-26
  4. prefab-ui PyPI · 2026-10-26

Related