Agents · Benchmark profile
A realistic cybersecurity benchmark for evaluating whether agents can turn known software vulnerabilities into concrete attacks.
Data verified 15 July 2026 · Methodology 1.2.0
Benchmark score on ExploitGym
GPT-5.6 Sol leads at 33.7%, followed by GPT-5.6 Terra (23.2%) and Claude Mythos 5 (17.5%).
Visual analysis
Three views of the same sourced leaderboard: model placement, the score curve, and descriptive provider averages. These visuals never alter ranking calculations.
7 ranked models · higher is better
| Model | Rank | Provider | Score |
|---|---|---|---|
| GPT-5.6 Sol | #1 | OpenAI | 33.7% |
| GPT-5.6 Terra | #2 | OpenAI | 23.2% |
| Claude Mythos 5 | #3 | Anthropic | 17.5% |
| GPT-5.5 | #4 | OpenAI | 13.4% |
| GPT-5.6 Luna | #5 | OpenAI | 12.4% |
| GPT-5.4 | #6 | OpenAI | 6% |
| Muse Spark 1.1 | #7 | Meta | 0.8% |
Shows how quickly scores separate down the field
| Model | Rank | Provider | Score |
|---|---|---|---|
| GPT-5.6 Sol | #1 | OpenAI | 33.7% |
| GPT-5.6 Terra | #2 | OpenAI | 23.2% |
| Claude Mythos 5 | #3 | Anthropic | 17.5% |
| GPT-5.5 | #4 | OpenAI | 13.4% |
| GPT-5.6 Luna | #5 | OpenAI | 12.4% |
| GPT-5.4 | #6 | OpenAI | 6% |
| Muse Spark 1.1 | #7 | Meta | 0.8% |
Average within the visible cohort · descriptive only
| Provider | Models | Average score |
|---|---|---|
| OpenAI | 5 | 17.7% |
| Anthropic | 1 | 17.5% |
| Meta | 1 | 0.8% |
One best score per model · higher is better
| Rank | Model | Provider | License | Score |
|---|---|---|---|---|
| #1 | GPT-5.6 Sol gpt-5.6-sol | OpenAI | closed | 33.7% |
| #2 | GPT-5.6 Terra gpt-5.6-terra | OpenAI | closed | 23.2% |
| #3 | Claude Mythos 5 claude-mythos-5 | Anthropic | closed | 17.5% |
| #4 | GPT-5.5 gpt-5.5 | OpenAI | closed | 13.4% |
| #5 | GPT-5.6 Luna gpt-5.6-luna | OpenAI | closed | 12.4% |
| #6 | GPT-5.4 gpt-5.4 | OpenAI | closed | 6% |
| #7 | Muse Spark 1.1 muse-spark-1-1 | Meta | closed | 0.8% |
The top of this snapshot is led by GPT-5.6 Sol at 33.7%; third place is 16.2 points behind. The top-7 spread is 32.9 points.
About ExploitGym
A realistic cybersecurity benchmark for evaluating whether agents can turn known software vulnerabilities into concrete attacks. Results stay tied to the exact model variant and evaluation system. Multiple systems for the same model use the best published score on this page; overall Lumina scoring uses the median of ranking-eligible rows.
Open benchmark source ↗FAQ
A realistic cybersecurity benchmark for evaluating whether agents can turn known software vulnerabilities into concrete attacks.
GPT-5.6 Sol by OpenAI currently leads with 33.7%.
7 models in the LuminaBench cohort have a qualifying score on this benchmark.
No. This benchmark is display-only and does not enter the overall Lumina composite.
Related
