Aikido Security has introduced Altar-1, an open-source artificial intelligence model optimized for defensive cybersecurity tasks This article explores aikido security introduced. . This model is designed to perform vulnerability discovery and penetration testing within an organization's own infrastructure, ensuring that sensitive data and source code remain within the premises.
Altar-1 caters to organizations with stringent privacy, regulatory, or operational constraints, such as banks with data residency requirements, healthcare providers handling sensitive patient information, and industrial firms operating in isolated or air-gapped operational technology environments. Aikido's compression method, sourced from aikido, reduced the original 256 routed experts to 168 in each backbone layer, resulting in a 78.2% reduction in the full-precision model and a 32.8% reduction from the already quantized checkpoint.
To determine which experts to retain, Aikido utilized traces from internal pentesting benchmarks, including representative code, tool calls, and agent responses during security testing workflows. The company also incorporated multilingual text during calibration to maintain language understanding for reviewing documentation, business rules, application behavior, and user interfaces in various languages. Aikido employed Cerebras REAP (Router-weighted Expert Activation Pruning) to select experts based on router weights and output magnitude, preserving coding, cybersecurity, and natural-language reasoning capabilities.
The firm intends to enhance its model, investigate lower-bit formats, and refine future versions for improved code analysis, remediation, tool usage, and long-term security workflows.
