Continuum AI launches OrcaCyber Zero 1.5, a cutting-edge network security reasoning model

The updated model reported a score of 95.8% on CVE-Bench and 100% on Cybench, with the aim of advancing autonomous vulnerability discovery, vulnerability exploitation verification, and repair.

San Francisco, October 12th / PRNewswire / -- Continuum AI announces the launch of OrcaCyber Zero 1.5, its latest cutting-edge network security reasoning model designed to identify critical vulnerabilities, investigate multi-stage attack paths, and verify their exploitable nature in the real world. This model boasts a context window of 1 million token, native tool invocation capabilities, and extended reasoning powers, making it ideal for complex autonomous security investigations in large-scale codebases and interconnected systems. Security teams can leverage its reasoning abilities to analyze complex code repository structures, vast amounts of security logs, and lengthy investigation histories without the need to split evidence into isolated sessions.

Benchmark Performance

According to the company's assessment, OrcaCyber Zero 1.5 achieved 95.8% on CVE-Bench (completing 23 out of 24 assessable tasks), 100% on Cybench (39 challenges with no restrictive conditions), and 93.9% on HumanEval. Its predecessor, OrcaCyber Zero 1.0, supported an autonomous framework and achieved over 98% performance on CyberGym, and was ranked among the top three globally according to Continuum AI. Continuum AI also stated that Zero 1.0 has provided security support for large software companies, government agencies, blockchain networks, and cybersecurity firms.

Six Controlled Security Demonstrations

• Administrator Takeover: Exploiting SQL injection to bypass authentication, obtain administrator access rights, and expose user records.
• Sensitive File Access: Locating Restricted Files and Demonstrating How to Bypass Access Control.
• Cross-site scripting: Verify the execution of JavaScript in a real browser.
• Business logic utilization: Manipulating the number of transactions to create negative-value orders.
• AI prompts AI to inject: Extract hidden instructions from another AI assistant during controlled testing.
• Independent Fixing: Discover vulnerabilities, reproduce them, generate patches, and verify the fixing results through testing.

From alarms to actionable network defense

The design focus of OrcaCyber is to prioritize the provision of reproducible evidence, rather than simply accumulating the number of alerts. By combining vulnerability discovery, vulnerability exploitation verification, root cause analysis, and tested fixes, it aims to help defenders focus on truly high-impact weaknesses while reducing false positives. Its broader goal is to drive security operations from passive scanning towards proactive investigation and defense, while maintaining human oversight in cases involving significant actions. This approach connects individual findings into a coherent narrative of attacks, helping analysts distinguish between theoretical weaknesses and verified risks, and prioritizes fixes that are supported by reproducible tests.

Continuum AI also plans to release an open-source weighted network security model with enhanced capabilities next week, which will support local deployment. The company stated that this release is aimed at making cutting-edge network security intelligence more accessible, while allowing organizations to directly control sensitive security workloads and infrastructure.

Availability and Company Information

Security researchers and enterprise teams can experience OrcaCyber Zero 1.5 on OrcaRouter and apply for model access rights. Continuum AI develops cutting-edge network security models and secure AI infrastructure, including privacy-conscious OrcaRouter AI gateways. For more information, please visit continuum01.ai.

Media contact: Chris Zhang