automation & business tech ·

Aikido ships Altar, an open-weight AI security model that runs fully local

Belgian cybersecurity startup Aikido released Altar on Monday, September 21 — an open-weight AI model built specifically for defensive cybersecurity, designed to be installed and run entirely inside customer-controlled systems, including air-gapped networks, so sensitive code never goes to an external provider. Altar is a compressed, customized version of Z.AI's open GLM-5.3 model, slimmed down through quantisation and expert pruning for local deployment; Aikido says the model is available openly for local use.

The driver is on both sides of the threat curve: rising concern that criminals are using AI to find and exploit software vulnerabilities is fueling demand for defensive AI tools that can run locally, and regulated firms are refusing to send their codebases to outside clouds at all. Altar builds on Aikido's autonomous penetration-testing tool, Aikido Machine, for code vulnerability detection and validation with full data residency. Aikido says it compressed the model from 1.51TB to 328GB, deploys it on a 4-H200s node with vLLM, and tested it against 32 known vulnerabilities across 30 repositories, retaining what the company says is 92% of the parent model's coverage — company claims, not independently measured. A critical-severity vulnerability was reportedly caught during a client test.

Reuters names Belgian bank Belfius — already a confirmed Aikido customer deploying its on-prem Aikido Machine with 200+ AI agents in its own data center — as one of the customers whose systems will get Altar. Aikido reached a $1 billion valuation in January 2026, and "sovereign AI security" is fast becoming its own category: the AI, without the data exposure.

Why it matters

Every small business scanning code for vulnerabilities faces the same tradeoff: use powerful cloud AI and hand over your code, or keep it local and use weaker tools. Altar is pitched as the end of that tradeoff — a purpose-built, locally-runnable defensive model. If the pattern holds, expect compressed derivatives of open bases tuned for regulated verticals to become a repeatable playbook.

Key facts

  • Released Monday, September 21, 2026 by Belgian firm Aikido (valued at $1B in January)
  • Open-weight (not open-source): compressed, customized version of Z.AI's open GLM-5.3, available openly for local deployments
  • Runs fully inside customer-controlled systems, including air-gapped networks — sensitive code never leaves the premises
  • Built on Aikido's Aikido Machine autonomous pentesting tool; Belgian bank Belfius named by Reuters as a customer whose systems will get Altar
  • Aikido claims: 1.51TB compressed to 328GB, deploys on a 4-H200s node, 92% of parent-model vulnerability coverage retained (company claims, not independently measured)

Sources

Spot an error? We correct quickly and note it. Contact us via the contact page.

← More AI news