Aikido Security introduced Altar-1 on September 21, its first open-weight model for defensive security. It is designed to run vulnerability analysis inside customer-controlled infrastructure, including air-gapped networks, without sending source code, architecture details, or findings to an external inference provider. The model powers Aikido Machine, the company’s autonomous penetration-testing appliance.

### A compressed GLM-5.3 derivative

Altar-1 was not trained from scratch. Aikido started with Z.AI’s GLM-5.3 mixture-of-experts model and combined quantization with expert pruning. The company reports that the original BF16 model occupies 1,506.7 GB, an AWQ INT4 checkpoint takes 488.2 GB, and Altar-1 reduces storage to 328 GB. It retains 168 of 256 experts per layer while removing 88.

Aikido says the model can be served through vLLM or SGLang on a node with four Nvidia H200 GPUs. That requirement is beyond an ordinary workstation, but it targets organizations with private data centers and strict controls on where sensitive material can go. The weights and deployment instructions are public, although open weights do not mean unrestricted terms: Altar-1 inherits the GLM-5.3 license, including a security-review condition for certain very large model-as-a-service operators.

### Vendor-published results

Aikido evaluated the model against 32 known vulnerabilities across 30 repositories, running each case three times. Altar-1 averaged 60.4% recall per run and found 23 of 32 vulnerabilities at least once. The unpruned quantized checkpoint averaged 61.5% and also found 23, while the full-precision parent averaged 65.6% and covered 25.

The numbers suggest that pruning preserved much of the parent model’s capability, but the benchmark was designed and run by the vendor. Independent testing is still needed to assess false positives, reproducibility, agent security, operating cost, and performance in different enterprise environments. The results do not establish that the model can replace human review, technical validation, or access controls.

The release illustrates a practical direction for AI security: moving the model to sensitive data instead of moving code and findings to an external cloud. Regulated organizations may gain operational control, but they also assume responsibility for hardware, updates, monitoring, and model governance. Altar-1 provides a testable option for that tradeoff while independent production evidence remains limited.

— Lía Torres / AI News