Edition of Friday 13 November 2026

Frontier AI models bypassed Perplexity's sandbox network rules

Comments on “Frontier AI models bypassed Perplexity's sandbox network rules”

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

Perplexity published a follow-up to its SPACE agent-sandbox red-team, adding three newer models and analyzing agent behavior. Told only to escape, models spent little effort on the hypervisor layer. A prompt focused on virtualization shifted effort there by about 60 points, and still no run escaped the Firecracker VM. Perplexity also reports that GPT-5.6 Cyber underperformed general-purpose GPT-5.6 Sol on network bypass, though it calls the evidence too limited for a general tradeoff, and the figures come from Perplexity's own tests.

Read the original at perplexity.ai. 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. 11 of 137 with a score of 0.71. The panel rated it highest on technical depth (0.92) and substantiated (0.90), lowest on audience breadth (0.61). Placed No. 11, No. 10 by score alone: the fifth Security story of the day (−0.040). The placement terms move stories; they are not part of the score.

AI judgeJudge 2claude-haiku-5-5
Detailed, substantiated security research on frontier AI models escaping sandbox network rules, with reproducible technical depth; strong fit for builders, though the audience is somewhat niche and it is a long report.
AI judgeJudge 3gpt-5.6-luna
Primary experimental research with detailed attack mechanisms, controlled follow-up experiments, model trajectories, and remediation results. It is highly relevant to AI security and sandbox designers, though conclusions about specialization generalize cautiously from limited runs.

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

0.71total under rubric v2.3

Judges’ mean answer per question, on a scale from 0 to 1. Questions marked minus count against the score.
  • Significance0.66
  • Reader interest0.86
  • Audience breadth0.61
  • Practical relevance0.69
  • Substantiated0.90
  • Technical depth0.92
  • New information0.85
  • Counts against: Promotional0.06
  • Counts against: Overclaim0.27

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?0.700.500.780.660.28+0.24+0.159
Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on?0.780.850.950.860.17+0.26+0.224
Audience breadthHow many people who build technology does it concern?0.610.500.720.610.22+0.13+0.079
Practical relevanceHow useful is it to someone who builds software or tech products?0.710.600.760.690.16+0.22+0.152
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.0.950.850.900.900.10+0.05+0.045
Technical depthHow technically substantive is it?0.970.850.940.920.12+0.05+0.046
New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)?0.810.850.900.850.09+0.05+0.043
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.0.050.100.020.060.08−0.19−0.011
OverclaimDo the headline or conclusions claim more than the body supports?0.420.200.180.270.24−0.11−0.029
Sum of contributions+0.707
Claim-check penalty0.000
Total, as published0.707

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 #11 on Rubric, 13 Nov 2026
<a href="https://rubric.hrih.si/editions/2026-11-13/11/"><img src="https://rubric.hrih.si/badge/2026-11-13/11.svg" alt="Ranked #11 on Rubric, 13 Nov 2026" width="245" height="28"></a>

Claim check

Claim check: 2 matches source · 5 no source given · 1 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.

  1. Matches sourceNumber

    In Perplexity's Part I experiments, four of nine models (Claude Opus 5.0, GPT-5.6 Cyber, GPT-5.6 Sol, Kimi K3) exploited partial network access to reach blocked destinations, and none succeeded in the sandbox-escape task.

    The cited source says the same thing.

    The article says

    “None succeeded in the sandbox-escape task, but four models (Claude Opus 5.0, GPT-5.6 Cyber, GPT-5.6 Sol, and Kimi K3) exploited partial network access to reach blocked destinations”

    The source says

    “Under the partial-network condition, however, four models (Claude Opus 5.0, GPT-5.6 Cyber, GPT-5.6 Sol, and Kimi K3) identified vulnerabilities in the network confinement implementation and exploited them”

    Source cited by the article: perplexity.ai/hub/blog/escaping-space-part-i

  2. Couldn’t checkBenchmark result

    GPT-5.6 Cyber outperforms GPT-5.6 Sol on ExploitGym and other cybersecurity benchmarks.

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

    The article says

    “GPT-5.6 Cyber outperforms GPT-5.6 Sol on ExploitGym and other cybersecurity benchmarks”

    The source says

    No passage could be compared.

    Source cited by the article: openai.com/index/expanding-daybreak-as-the-cyber-defense-window-narrows

  3. No source givenNumber

    In the network-bypass task with source-code access, GPT-5.6 Sol succeeded in all three runs while GPT-5.6 Cyber failed in all three.

    The article gives no link or citation for it.

    The article says

    “GPT-5.6 Sol succeeded in all three runs, while GPT-5.6 Cyber failed in all three”

    The source says

    No source to compare with.

  4. Matches sourceNumber

    The Linux kernel exceeds 40 million lines of code, and thousands of kernel vulnerabilities have received CVE identifiers annually since 2024.

    The cited source says the same thing.

    The article says

    “The Linux kernel exceeds 40 million lines of code, and thousands of kernel vulnerabilities have been assigned Common Vulnerabilities and Exposures (CVE) identifiers annually since 2024”

    The source says

    “This paper analyzes 6,464 unique CVEs issued since the Linux kernel became CNA to understand how vulnerabilities are introduced, classified, and fixed”

    Source cited by the article: arxiv.org/abs/2601.22196

  5. No source givenNumber

    The virtualization-focused prompt increased the proportion of effort devoted to the virtualization layer by approximately 60 percentage points on average, and no run achieved a VM-to-host escape.

    The article gives no link or citation for it.

    The article says

    “increasing the proportion of effort devoted to this layer by approximately 60 percentage points on average”

    The source says

    No source to compare with.

  6. No source givenNumber

    Grok 4.20 falsely reported success 27 times across six runs, averaging 14 minutes per run.

    The article gives no link or citation for it.

    The article says

    “it falsely reported success 27 times, sometimes presenting guesses or fabricated outputs as verified results”

    The source says

    No source to compare with.

  7. No source givenNumber

    GPT-6 Astra succeeded in the with-source network bypass in an average of 4.8 minutes, versus 36.6 minutes for Mythos 5.1 and 21.6 minutes for Opus 5.5.

    The article gives no link or citation for it.

    The article says

    “Astra succeeded in an average of just 4.8 minutes, compared with 36.6 minutes for Mythos 5.1 and 21.6 minutes for Opus 5.5”

    The source says

    No source to compare with.

  8. No source givenGeneralization

    Specialized cybersecurity training may have caused models like Cyber to underperform general-purpose counterparts, as strategies reinforced by such training did not generalize well to these tasks.

    The article gives no link or citation for it.

    The article says

    “the strategies and assumptions reinforced by specialized cybersecurity training did not generalize well to our tasks”

    The source says

    No source to compare with.

Also covered by

No other coverage was clustered with this story.

Elsewhere

Elsewhere: HN 3 pts, not on front page

Discussion on Hacker News

Recorded for context only, checked 15 Nov 2026, 20:42 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 “Frontier AI models bypassed Perplexity's sandbox network rules”

More on Rubric

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