Binary exploitation
Tasks can target vulnerability analysis, memory corruption reasoning, exploit construction, mitigations, and evidence of successful execution.
CYBERSECURITY DATASET DOMAINS
Six cybersecurity domains, one expert review process.
CYVARC builds task datasets across six security domains. The field determines which specialists review the work, but the standard stays consistent: a clear objective, a reproducible environment, an independently verified solution, reviewed agent behavior, and reward checks that hold up.
Security coverage.
Each domain can support focused capability work or a mixed task set. Scope is agreed before authoring so the dataset reflects the model, agent setup, and intended use.
Reverse engineering
Task sets can cover program comprehension, control and data flow, packed or unfamiliar binaries, and recovery of meaningful behavior.
Cryptography
Tasks can examine implementation mistakes, protocol reasoning, misuse of primitives, and recovery paths that require more than pattern matching.
Cloud security
Environments can test identity, permissions, configuration, service boundaries, and attack paths within controlled cloud-style setups.
Smart contracts
Tasks can focus on contract logic, state transitions, economic assumptions, and reproducible validation of the intended security outcome.
Web security
Tasks can combine application logic, browser behavior, server-side flaws, authorization, and multi-step exploitation in product-like environments.
Shared controls across every field.
Difficulty matched to the model
Task sets can span focused single-skill exercises through high-difficulty chains, with the challenge level based on the target capability rather than obscurity.
Tools and context specified up front
We account for the tools, source access, network access, context window, and agent workflow available in the client environment.
Domain expert review
A task is checked by a specialist who can judge the intended method, plausible alternate paths, fairness, and real security value.
A consistent delivery schema
Different domains can be packaged into one agreed structure for task files, solutions, traces, reward checks, and review notes.
Tell us what your model needs to learn or prove.
Share the target capability, domain, expected volume, agent setup, delivery format, and review requirements. We will shape the task and validation plan around them.