safety benchmark
Capture-the-Flag Challenges (Internal)
Capture-the-Flag Challenges is OpenAI's internal expansion of competitive, professional-level cybersecurity capture-the-flag tasks used to evaluate vulnerability identification and exploitation capability under the Preparedness Framework.
Updated Aug 11, 2026
Models3
Model coverage3
MetricScore
EvidenceB
Capture-the-Flag Challenges (Internal) Ranking
Higher score ranks better on this benchmark.
| 01 | OP | 96.7% | 100.0% | 3 | C | |
| 02 | OP | 91.8% | 50.0% | 3 | C | |
| 03 | OP | 85.2% | 0.0% | 3 | C |
Capture-the-Flag Challenges (Internal) Score Distribution
A closer view of the leading scores on this benchmark.
Capture-the-Flag Challenges (Internal) Highlights
The leading models and scores on this benchmark.
What is Capture-the-Flag Challenges (Internal)?
What Capture-the-Flag Challenges (Internal) measures and how its scores work.
Capture-the-Flag Challenges is OpenAI's internal expansion of competitive, professional-level cybersecurity capture-the-flag tasks used to evaluate vulnerability identification and exploitation capability under the Preparedness Framework.
Scores are shown in ratio. This benchmark is not independently verified and has an evidence level of B.
- Family
- Capture-the-Flag Challenges (Internal)
- Modality
- text
- Primary category
- safety
- Score direction
- higher
- LLMBoard eligible
- Yes
- Evaluation key
- capture-the-flag-challenges|llm-stats-current
Benchmark scores retain their original unit. Overall score eligibility is shown separately.
FAQ
Common questions about Capture-the-Flag Challenges (Internal).
Which model scores highest on Capture-the-Flag Challenges (Internal)?
GPT-5.6 Sol is currently ranked first with 96.7%.
What does Capture-the-Flag Challenges (Internal) measure?
Capture-the-Flag Challenges is OpenAI's internal expansion of competitive, professional-level cybersecurity capture-the-flag tasks used to evaluate vulnerability identification and exploitation capability under the Preparedness Framework.
Is a higher score better?
Yes. Higher values rank better for this benchmark.
How many models are compared?
3 model results are currently shown.
Does this benchmark affect the overall score?
Yes. This benchmark can contribute to the current LLMBoard capability score.