AlgoProof AI Safety Certification Framework Earns ISO Standard: 127 Tests Required Before Market Entry

AlgoProof AI Safety Certification Framework Earns ISO Standard

On October 3, 2028, the International Organization for Standardization (ISO) officially released ISO 23891, incorporating the AlgoProof algorithm safety certification framework into the global AI product admission system. Starting January 2029, autonomous driving systems, medical AI diagnostic tools, and financial trading algorithms sold in the EU and Japanese markets must pass AlgoProof's 127 safety tests.

AlgoProof was initiated in 2026 by the Algorithm Safety Foundation (ASF), a Zurich-based nonprofit, and took two years to complete standardization. Its tests cover three dimensions: technical robustness (42 tests), fairness and bias detection (38 tests), and explainability and transparency (47 tests).

Over the past two years, ASF conducted pre-certification testing on 3,400 AI systems globally, finding that 38% exhibited serious discriminatory bias in fairness tests, such as systematically lower credit scores for specific ethnic groups. These systems previously had no mechanism to surface such issues.

The AlgoProof certification process consists of three stages. The first is automated static analysis, machine-reviewing training data, architecture design, and output patterns, taking approximately 48 hours. The second is red team adversarial testing, where certified security researchers attack the system to find exploitable vulnerabilities. The third is social impact assessment, where interdisciplinary expert panels review potential societal effects in real-world scenarios.

EU Digital Affairs Commissioner Henna Virkkunen stated that AlgoProof will become one of the designated certification frameworks under the EU AI Act. She emphasized this is not about creating barriers to innovation, but ensuring AI products entering the market are safe, fair, and explainable. Consumers have the right to know the products they use have undergone rigorous safety review.

Certification costs are a key industry concern. ASF says a complete certification costs between $150,000 and $800,000 depending on system complexity. While manageable for large tech companies, this may burden startups. ASF established a Fast Track program offering subsidized certification at $30,000 for businesses with annual revenue below $10 million.

Notably, AlgoProof currently covers only three high-risk sectors. ASF plans to expand to education AI, human resources AI, and public safety AI in 2029, with full industry coverage by 2030.

Critics argue AlgoProof is overly conservative and may hinder AI innovation. MIT AI Lab professor David Park points out that many of the 127 tests were designed based on 2026 risk understanding, and AI technology evolves far faster than standards update. The standard may already be outdated by release.

ASF responds that AlgoProof uses a living standard mechanism, updating test items quarterly based on newly discovered risks, rather than the traditional multi-year revision cycle. Standards should not become obstacles to innovation, but innovation should not become an excuse for safety.