Checked Nov 13, 2026Start a new check

What was checked

OpenAI presented an unreleased cybersecurity model scoring 77% on ExploitGym, compared with 52.6% for OpenAI GPT-6 Cyber, 54.8% for Google Gemini 4 Carbon and 39.2% for Anthropic Claude Mythos 5.2.

Assessment summary

Needs caution

OpenAI's published system materials list GPT-6 Cyber at about 48.5% on ExploitGym, while some live-reporting of the keynote referenced a ~52% figure; the 52.6% number does not match the formal system card value and likely mixes a live remark with other measurements. Secondary reports describe a livestream or leaked chart showing a 77.0% score, but they also note the model had no name, system card, or external review.

Key findings

Model assessments

Assessment summary

The 77% result is reported, but only through secondary and unverified coverage rather than an authoritative OpenAI publication.tokenstrom.com finance.biggo.com The comparison figures are inconsistent across available sources: one source lists 52.6% for GPT-6 Cyber and 54.8% for Gemini 4 Carbon, while another reports 48.45% and 53.5%, respectively.benchgap.net innovaiden.com

Claims reviewed

GPT-6 Cyber scored 52.6% on ExploitGym

Can’t verify
Some secondary reports give 52.6%, but another leaderboard reports 48.5% and another source cites OpenAI's system-card result as 48.45%.benchgap.net agihunt.info innovaiden.com

Google Gemini 4 Carbon scored 54.8% on ExploitGym

Can’t verify
One source attributes 54.8% to Google, while another leaderboard reports 53.5%; the available material does not establish which testing conditions or figure is authoritative.benchgap.net innovaiden.com

Anthropic Claude Mythos 5.2 scored 39.2% on ExploitGym

Can’t verify
Available sources report approximately 38% or 38.1%, not 39.2%, but the evidence is secondary and does not fully document the testing methodology.finance.biggo.com benchgap.net

Assessment summary

  • The claim contains multiple significant factual errors in the reported benchmark scores.
  • GPT-6 Cyber's actual ExploitGym score is 48.5%, not 52.6%.
  • Claude Mythos 5.2 scores approximately 38%, not 39.2%.
  • The 77% score for an unreleased model is not clearly documented as an ExploitGym result.

Claims reviewed

39.2% for Anthropic Claude Mythos 5.2

Contradicted
Multiple sources confirm Claude Mythos 5.2 scores approximately 38% or 38.1% on ExploitGym, not 39.2%.

This claim may mislead readers without added context.

52.6% for OpenAI GPT-6 Cyber

Contradicted
OpenAI's official launch post (November 5, 2026) reports GPT-6 Cyber scores 48.5% on ExploitGym. The November 12 keynote mentioned 52%, but the authoritative figure is 48.5%.

This claim may mislead readers without added context.

54.8% for Google Gemini 4 Carbon

Needs context
Google reported 54.8% for Gemini 4 Carbon with a ten-hour cap per task. However, the standard ExploitGym score cited by most sources is 53.5%, making this figure technically accurate but using a different testing condition.

This claim may mislead readers without added context.

Assessment summary

  • Some of the numeric claims are supported by public reporting, but sources disagree on the exact GPT-6 Cyber figure and the "unreleased" 77% score is only shown briefly in a presentation and not documented in a formal system card.
  • Because reporting and provider system cards differ, the overall statement mixes verified numbers with a partially corroborated teaser.

Claims reviewed

OpenAI GPT-6 Cyber scored 52.6% on ExploitGym

Needs context
OpenAI's published system materials list GPT-6 Cyber at about 48.5% on ExploitGym, while some live-reporting of the keynote referenced a ~52% figure; the 52.6% number does not match the formal system card value and likely mixes a live remark with other measurements deploymentsafety.openai.com fourweekmba.com.

This claim may mislead readers without added context.

OpenAI presented an unreleased cybersecurity model scoring 77% on ExploitGym

Supported
A briefly shown slide from OpenAI's November 12 presentation placed an "upcoming/internal" cyber model at 77.0% on ExploitGym, and media coverage captured that chart explainx.ai thefrontierwire.com.

Anthropic Claude Mythos 5.2 scored 39.2% on ExploitGym

Supported
Reporting and Anthropic materials place Mythos 5.2 in the high-30% range on ExploitGym; some summaries compute ~39.2% (others report ~38%), so 39.2% is within reported values assets.claude.com benchgap.net.

Sources cited

20 sources

Share this result