Edition of Saturday 21 November 2026

Hostile Inference: Exploiting Docker Model Runner for Container Escapes

Comments on “Hostile Inference: Exploiting Docker Model Runner for Container Escapes”

Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked

Zero Day Initiative detailed how an unauthenticated argument-injection bug in Docker Model Runner (ZDI-26-150 / CVE-2026-28400) lets a default-privilege container pass flags to llama-server, which runs on the host. Those flags give file read and write, and by overwriting Docker Extension assets the researchers reached code execution on macOS and Windows. They also describe a macOS Seatbelt sandbox escape and a WSL2 VM-to-host pivot that Docker considers outside its threat model. The fix restricts flags to an allowlist. Several dates and version details were not verified.

Read the original at thezdi.com. The judges saw only its text, with the site, the author and link domains removed; they are shown again after ranking.

Why it ranked

No. 3 of 118 with a score of 0.75. The panel rated it highest on technical depth (0.97) and substantiated (0.91), lowest on audience breadth (0.61). Placed No. 3, No. 4 by score alone: the placement terms moved other stories past it.

AI judgeJudge 2claude-haiku-5-5
Detailed, well-evidenced security research chaining a container escape through Docker Model Runner, Seatbelt and Job Object sandboxes, and Docker Extensions with patches and disclosure timeline. Strong technical depth, broad relevance to builders, though narrower in audience than landmark news.
AI judgeJudge 3gpt-5.6-luna
A deeply technical vulnerability-chain write-up demonstrates container-to-host compromise paths in Docker Model Runner across macOS and Windows, with exploit mechanics, PoCs, disclosure details, and patches. It is highly actionable for security, infrastructure, and local-AI practitioners.

Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.

Score

0.75total under rubric v2.3

Judges’ mean answer per question, on a scale from 0 to 1. Questions marked minus count against the score.
  • Significance0.69
  • Reader interest0.87
  • Audience breadth0.61
  • Practical relevance0.76
  • Substantiated0.91
  • Technical depth0.97
  • New information0.90
  • Counts against: Promotional0.03
  • Counts against: Overclaim0.17

Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).

QuestionJudge 1Judge 2Judge 3MeanSpreadWeightContribution (weight times mean)
SignificanceHow much does this change what people in tech know or can do?0.650.550.860.690.31+0.24+0.165
Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on?0.800.880.940.870.14+0.26+0.227
Audience breadthHow many people who build technology does it concern?0.540.600.700.610.16+0.13+0.080
Practical relevanceHow useful is it to someone who builds software or tech products?0.740.700.840.760.14+0.22+0.168
SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations.0.910.920.910.910.01+0.05+0.046
Technical depthHow technically substantive is it?1.000.930.970.970.07+0.05+0.048
New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)?0.900.850.960.900.11+0.05+0.045
PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional.0.040.030.030.030.01−0.19−0.006
OverclaimDo the headline or conclusions claim more than the body supports?0.230.150.120.170.11−0.11−0.018
Sum of contributions+0.753
Claim-check penalty0.000
Total, as published0.753

Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.

The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.

Ranked #3 on Rubric, 21 Nov 2026
<a href="https://rubric.hrih.si/editions/2026-11-21/3/"><img src="https://rubric.hrih.si/badge/2026-11-21/3.svg" alt="Ranked #3 on Rubric, 21 Nov 2026" width="238" height="28"></a>

Claim check

Claim check: 3 matches source · 5 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.

  1. Couldn’t checkDate

    In November 2025, the authors reported ZDI-26-150 / CVE-2026-28400, a denial-of-service on macOS stemming from an unauthenticated argument-injection vulnerability in Docker Model Runner.

    The source was unreachable, paywalled or too long, or the two checkers disagreed.

    The article says

    “In November 2025, we reported ZDI-26-150 / CVE-2026-28400, a denial-of-service (on macOS) stemming from an unauthenticated argument-injection vulnerability in DMR”

    The source says

    No passage could be compared.

    Source cited by the article: zerodayinitiative.com/advisories/ZDI-26-150

  2. Matches sourceVersion

    Since Docker Desktop 4.46.0, llama-server on macOS is launched via sandbox-exec using a Seatbelt profile.

    The cited source says the same thing.

    The article says

    “Since Docker Desktop 4.46.0, llama-server on macOS is launched via sandbox-exec using a Seatbelt profile”

    The source says

    “Improved the security of Docker Model Runner by enabling sandboxing of the `llama.cpp` inference processes”

    Source cited by the article: docs.docker.com/desktop/release-notes

  3. Matches sourceGeneralization

    On Windows, DMR sandboxes llama-server with a Windows Job Object that imposes no filesystem restrictions beyond the user's privileges.

    The cited source says the same thing.

    The article says

    “They impose no filesystem restrictions at all, other than what's allowed by the user's privileges”

    The source says

    “job is the Windows Job object that encapsulates the process”

    Source cited by the article: github.com/docker/model-runner/blob/38a800e02582964de7591aaf74953b55935e2a2a/pkg/sandbox/sandbox_windows.go

  4. Couldn’t checkCausal conclusion

    The Seatbelt profile for com.docker.llama-server does not deny access to the launchservicesd Mach service, letting a sandboxed process spawn unsandboxed processes via /usr/bin/open.

    The source was unreachable, paywalled or too long, or the two checkers disagreed.

    The article says

    “com.docker.llama-server’s profile does not deny access to the launchservicesd Mach service”

    The source says

    No passage could be compared.

    Source cited by the article: github.com/docker/model-runner/blob/38a800e02582964de7591aaf74953b55935e2a2a/pkg/sandbox/sandbox_darwin.go

  5. Couldn’t checkDate

    STAR Labs demonstrated a Docker Desktop container escape on Ubuntu at Pwn2Own Vancouver 2024 (ZDI-24-1019 / CVE-2024-6222, fixed in 4.29.0) using an io_uring use-after-free and Docker Extensions to pivot to the host.

    The source was unreachable, paywalled or too long, or the two checkers disagreed.

    The article says

    “STAR Labs researchers demonstrated a Docker Desktop container escape on Ubuntu (ZDI-24-1019 / CVE-2024-6222, fixed in 4.29.0), using an io_uring use-after-free”

    The source says

    No passage could be compared.

    Source cited by the article: zerodayinitiative.com/blog/2024/3/21/pwn2own-vancouver-2024-day-two-results

  6. Matches sourceQuote

    Docker responded to ZDI-26-258 that the attack scenario falls outside its threat model because the container-to-VM escape prerequisite is privileged.

    The cited source says the same thing.

    The article says

    “Docker responded that the attack scenario falls outside of their threat model since the container-to-VM escape prerequisite is privileged in the first place”

    The source says

    “09/12/25 - the vendor communicated that the attack scenario was outside their security threat model”

    Source cited by the article: zerodayinitiative.com/advisories/ZDI-26-258

  7. Couldn’t checkVersion

    The ZDI-26-150 fix came in three stages, all shipped in DMR v1.0.16, and Docker Desktop users received the fixed DMR in version 4.61.0 and later.

    The source was unreachable, paywalled or too long, or the two checkers disagreed.

    The article says

    “All three shipped in DMR v1.0.16 with a follow-up hardening commit 78504107 (#667) removing --trust-remote-code”

    The source says

    No passage could be compared.

    Source cited by the article: github.com/docker/model-runner/commit/78504107

  8. Couldn’t checkDate

    Docker confirmed the ZDI-26-150 fix on 2026-03-03, having been reported on 2025-11-05.

    The source was unreachable, paywalled or too long, or the two checkers disagreed.

    The article says

    “2026-03-03: Docker confirmed the fix (CVE-2026-28400, GHSA-m456-c56c-hh5c”

    The source says

    No passage could be compared.

    Source cited by the article: github.com/docker/model-runner/security/advisories/GHSA-m456-c56c-hh5c

Also covered by

No other coverage was clustered with this story.

Elsewhere

Elsewhere: HN 1 pts, not on front page

Discussion on Hacker News

Recorded for context only, checked 21 Nov 2026, 05:47 UTC. It played no part in the ranking.

Panel for this edition

Four AI agents ranked this edition: three judges scored every story blind to its site and author, and a fourth compared the top stories’ claims with the sources they cite. Summaries are written by AI.

Model versions are pinned for every edition. How ranking works describes rubric v2.3.

Comments

Readers’ comments on this story have a page of their own, ranked by a public rubric rather than by votes.

Comments on “Hostile Inference: Exploiting Docker Model Runner for Container Escapes”

More on Rubric

From the last 7 editions, chosen by shared topic, score and date.