01 / AI · Claim vs. Check · Analysis
CTF-Dojo offers 658 tasks for language model training
CTF-Dojo provides verifiable execution results for agent training, while its published performance comparisons concern CTF security challenges.

Researchers released CTF-Dojo, an execution environment containing 658 security challenges for training language model agents.
The paper's abstract identifies a shortage of execution environments as a constraint on agent training. The research addresses that constraint through tasks whose execution results can be verified and used as feedback.
The researchers emphasised reproducible task execution and automated preparation of training environments. These underpin the training approach, while the published performance comparisons concern CTF security challenges.
Execution results as training signals
The abstract describes an environment in which agents can run tasks and receive verifiable feedback. This makes execution results available as training signals.
The researchers said they packaged the tasks in Docker containers, which bundle the conditions needed to run a program reproducibly. The tasks follow Capture-The-Flag (CTF), a format for solving security challenges.
According to the abstract, CTF-Forge automates the conversion of publicly available materials into ready-to-use execution environments. The researchers said it handles preparation in minutes that traditionally required weeks of expert configuration.
Performance on CTF security challenges
The researchers said they trained agents on 486 execution-verified trajectories, or sequences of steps taken through tasks. Their best-performing 32B model achieved a first-attempt success rate, known as Pass@1, of 31.9%, a researcher-published benchmark figure.
The researchers described that result as comparable to models including DeepSeek-V3-0324. The comparison in the abstract concerns evaluations on CTF security challenges.
The published results connect training on verified execution with performance on CTF tasks. The researchers did not disclose results for detecting vulnerabilities in live services in the abstract, leaving that application outside the reported comparison.