Aikido launches Altar-1 AI for local cyber defense

Belgian firm Aikido Security has introduced Altar-1, an AI model for vulnerability discovery and code analysis that can run on an organization’s local infrastructure.

Aikido says the model’s accessible weights and local deployment are intended to reduce reliance on cloud services and prevent sensitive code and security results from leaving internal networks.

Altar-1 was not trained from scratch. Aikido built it on GLM-5.3 from Z.AI and reduced the model’s size with quantization and expert pruning.

Aikido says full GLM-5.3 weighs about 1.51 terabytes, which quantization reduced to 488 gigabytes and pruning further cut to 328 gigabytes.

In pruning Aikido removed 88 of 256 expert sections, about 34.4 percent of the experts.

Aikido says Altar-1 is optimized for software vulnerability identification, code analysis, tool calling, reasoning and penetration-testing workflows, and can run on four H200 GPUs using vLLM.

The company highlights the model’s relevance for organizations that cannot allow code or security data to leave internal networks, notably in banking, healthcare and critical infrastructure environments.

For evaluation Aikido tested Altar-1 on a set of 32 known vulnerabilities across 30 code repositories. Altar-1 detected 60.4 percent of the vulnerabilities overall and found at least one instance of 23 of the 32 issues.

By comparison Aikido reports a quantized GLM-5.3 detection rate of 61.5 percent and the original GLM-5.3 at 65.6 percent. Aikido ran the test and says the results do not necessarily reflect performance on unknown vulnerabilities or in real-world conditions.

Aikido says Altar-1 is the first step of Aikido Labs’ plan to develop dedicated security models for vulnerability research, code analysis, automated fixes and advanced penetration testing. The model weights are published on Hugging Face and local deployment is supported.

Technical note: the Hugging Face page lists Altar-1 as roughly a 504 billion-parameter model, while the 328-gigabyte figure refers to the INT4 weight size; Aikido warns those two numbers are not equivalent.