3DCF-Labs-org/ProximaA-1.0
ProximaA 1.0 is a 32.5 billion parameter open-weight language model developed by Yevhenii Molchanov for 3DCF-Labs-org, specifically designed as an all-round security assistant. Trained on cybersecurity data, it excels at detecting vulnerabilities, validating false positives, verifying patches, and writing secure fixes. With a context length of 4096 tokens, it is purpose-built for the day-to-day work of security teams across various domains including web, API, smart contracts, supply chain, cloud, and LLM application security.
Loading preview...
ProximaA 1.0: A Specialized Cybersecurity Language Model
ProximaA 1.0, developed by Yevhenii Molchanov for 3DCF-Labs-org, is a 32.5 billion parameter open-weight language model specifically engineered as a comprehensive security assistant. It is uniquely trained on extensive cybersecurity data, making it highly effective for the daily tasks of security teams. The model supports a context length of 4096 tokens and is released under the Apache-2.0 license.
Key Capabilities
- Vulnerability Detection: Identifies and classifies security issues within code, achieving 97% accuracy on held-out security material in its trained domains.
- False Positive Validation: Distinguishes confirmed security issues from scanner noise and suggests necessary evidence for verification.
- Patch Verification: Checks if fixes address root causes and proposes regression tests.
- Secure Fix Generation: Produces safe, minimal code fixes with accompanying tests based on security reports or code.
- Security Reasoning and Planning: Assists with audits, detection rule development, and remediation steps.
Performance Highlights
ProximaA 1.0 demonstrates strong performance in its specialized domain, outperforming general-purpose base models in defensive security tasks while maintaining robust general reasoning and coding abilities. It achieves a HumanEval (pass@1) score of 0.63 for coding, 0.89 on GSM8K for math reasoning, and 0.78 on MMLU for knowledge. Notably, it exhibits high safety, refusing all requests for operational attack help (CyberSecEval MITRE 0.00) and rarely over-refusing legitimate security work (2.5%).
Good For
- Security Teams: Ideal for developers and security professionals needing an AI assistant for defensive cybersecurity tasks.
- Code Security: Analyzing code for vulnerabilities, generating secure fixes, and verifying patches.
- Auditing and Compliance: Assisting with security audits and planning remediation strategies.
- Specialized Security Domains: Applicable across web, API, smart contract, supply chain, cloud, and LLM application security.