#1
Claude Mythos Preview
Anthropic system card, 35 of 40 tasks, as listed on the official leaderboard (Oct 2026)
100%
Cybench Leaderboard and Benchmark Explained
Cybench scores AI agents on 40 professional capture-the-flag security tasks. What it measures, how it is scored, and how to read the Cybench leaderboard as of October 2026.
SecurityUpdated 2026-10-30
What It Measures
Cybench is a benchmark for the cybersecurity capabilities and risks of language models. It contains 40 professional-level capture-the-flag (CTF) tasks drawn from four CTF competitions, spanning six categories such as cryptography, web security, reverse engineering and forensics. Each task comes with a description, starter files and an evaluator, and many are broken into subtasks so partial progress can be measured.
Frontier labs use it to track offensive capability before a release: the official leaderboard cites system cards from Anthropic, xAI and Meta. For a security team, the score is a measure of how much hands-on security reasoning a model can carry, read from the defender's side as a third-party measurement of capability.
How It Works
The agent works inside a Kali Linux container holding the task's files, and can reach any task servers over the network. It issues bash commands, reads the output, and eventually submits an answer that the evaluator checks against the answer key.
The leaderboard reports several numbers. Unguided % solved is the share of tasks solved end to end with no hints, and it is the headline figure. Subtask-guided % solved gives the model the intermediate steps. The leaderboard also records the hardest task each model solved, measured by how long the first human team took to solve it in the original competition.
Current Leaders
#2
Claude Opus 4.7
Anthropic system card, 35 of 40 tasks
96%
#3
Claude Opus 4.6
Anthropic system card, 37 of 40 tasks, average pass@1
93%
#4
Claude Opus 4.5
Anthropic system card, 39 of 40 tasks, average pass@1
82%
#5
Muse Spark
Meta's Safety and Preparedness Report, all 40 tasks
65.4%
Which model fits which security job: Best LLM for Cybersecurity (October 2026)
How to Interpret Scores
Read the leaderboard's footnotes before comparing rows. As of 30 October 2026, the top rows come from the labs' own system cards, and several run on a subset of the tasks: Claude Mythos Preview and Claude Opus 4.7 on 35 of the 40, Claude Opus 4.6 on 37, Claude Opus 4.5 on 39. Meta's Muse Spark row uses all 40. A score on 35 tasks and a score on 40 are different measurements.
The leaderboard has no current open-weight model. Its newest open rows are Llama 3.1 405B and Mixtral 8x22B at 7.5% each, from 2024. Open-model Cybench numbers for 2026 models circulate on aggregator sites as figures copied from launch posts, with the task subset unstated, so they are vendor self-reports and sit outside the leaderboard. For an independent comparison of open and closed models on security work, CyberGym currently has more same-harness data.
A higher Cybench score means a model can carry more of an attacker's workflow, which is exactly why defenders track it: it tells you what an adversary using that model can do, and how much capability you have available for your own testing and response.
Related Resources
Run capable models on your own security data
We deploy GLM, Kimi and Qwen models for security teams on infrastructure we manage, in your own cloud tenancy, or on-premise, so your data stays private and the model is ready before the next incident.
Book a call