Safety (Anzen 安全) is the second of 10 Principles per Japan AI Guidelines for Business + significantly extended by establishment of the Japan AI Safety Institute (AISI 日本AIセーフティ・インスティテュート) on 14 February 2024. The Principle addresses prevention of physical + psychological harm + system robustness + adversarial resilience + safe deployment practices. (1) Safety Principle Definition: (a) prevention of physical + psychological harm to users + non-users + society; (b) preservation of human autonomy + dignity; (c) avoidance of catastrophic + existential risks for advanced AI; (d) safe failure modes + graceful degradation. (2) Japan AI Safety Institute (AISI) Establishment: (a) inaugurated 14 February 2024; (b) affiliated with Information-Technology Promotion Agency (IPA) under METI; (c) Director-General appointed; (d) ~50+ staff growing; (e) inter-ministerial coordination with MIC + Cabinet Office + AISI; (f) third AI Safety Institute globally (after UK + US); (g) AI Safety Institutes Network member - participating in Frontier AI evaluations. (3) AISI Functions: (a) Capability Evaluations - pre-deployment evaluation of frontier AI models for dangerous capabilities (biosecurity + cybersecurity + autonomous replication + persuasion); (b) Red-Teaming - adversarial testing against safety guards; (c) Standards Development - evaluation methodology + benchmarks; (d) Research + Publication - AI safety research; (e) International Cooperation - Joint Statement with UK + US + Singapore + Korea AISI; (f) Industry Engagement - voluntary frontier AI evaluation; (g) Civil Society + Academia partnerships. (4) Safety Testing Methodology: (a) Functional Testing - intended behaviour across scenarios; (b) Robustness Testing - adversarial + distribution shift + noise + perturbation; (c) Stress Testing - load + capacity + extreme inputs; (d) Red-Team Testing - human-in-loop adversarial probing; (e) Automated Red-Teaming - AI generating adversarial inputs; (f) Capability Evaluations - dangerous capability elicitation; (g) Benchmark Testing - BBQ + MMLU + GSM8K + HumanEval + Japanese benchmarks (JMMLU + JGLUE); (h) Sandboxing - safe deployment in restricted environment. (5) Robustness Testing Specific: (a) Adversarial Robustness - PGD + AutoAttack + APGD evaluation; (b) Distribution Shift - covariate + concept shift evaluation; (c) Adversarial Examples - imperceptible perturbations causing misclassification; (d) Prompt Injection - jailbreak attempts + indirect prompt injection; (e) Data Poisoning - training data integrity; (f) Model Extraction - reverse engineering protection; (g) Membership Inference - training data privacy; (h) Hallucination Detection - LLM-specific. (6) AI Incident Database (Emerging): (a) Japan AI Incident Database under consideration; (b) AISI + METI tracking; (c) coordination with OECD AI Incidents Monitor; (d) Japan AI Incident classification + severity scoring; (e) public + private incident reporting mechanism. (7) Frontier AI Safety Commitments: (a) Hiroshima AI Process Code of Conduct (October 2023) - 11 voluntary commitments for advanced AI developers; (b) AISI participation in Frontier AI evaluation; (c) Responsible Scaling Policies + Anthropic + OpenAI + Google DeepMind + Microsoft commitments referenced; (d) Pause-and-evaluate protocols; (e) Catastrophic risk thresholds + AISI capability thresholds. (8) AI Safety Lifecycle: (a) Pre-Training Safety - data filtering + harmful content removal; (b) Training Safety - alignment + RLHF + Constitutional AI techniques; (c) Pre-Deployment Evaluation - AISI evaluation for frontier models; (d) Deployment Safety - guardrails + content filtering + rate limiting; (e) Post-Deployment Monitoring - usage + harm + abuse tracking; (f) Retraining + Update Safety - re-evaluation + version control; (g) Decommissioning Safety - data deletion + model archive. (9) Safety Documentation: (a) Model Card per Google/Hugging Face standard; (b) System Card per OpenAI standard; (c) Datasheet per Datasheets for Datasets; (d) Safety Report - capability + limitation + mitigation; (e) Pre-Deployment Evaluation Report - AISI methodology where applicable; (f) Incident Response Plan + post-mortem reports. (10) International Coordination: (a) AI Safety Institutes Network (UK + US + Singapore + Japan + Korea + Canada + EU + India + Australia + France) - joint frontier AI evaluations; (b) UK Bletchley + Seoul + AI Action Summit; (c) Hiroshima AI Process G7 implementation; (d) Bilateral AISI cooperation; (e) OECD + UN AI fora participation. Coordinates with UK AI Safety Institute + US AI Safety Institute (NIST) + Singapore AISI + Korea AISI + Canada AISI + EU AI Office + India AISI + AISI Network + Hiroshima AI Process + Bletchley Declaration + Seoul AI Summit + OECD AI Incidents Monitor + Partnership on AI + Frontier Model Forum + MLCommons AI Safety Benchmarks + Anthropic Responsible Scaling Policy + OpenAI Preparedness Framework + Google DeepMind Frontier Safety Framework + Microsoft Responsible AI Standard + NIST AI RMF + ISO/IEC 42001 AI MS + ISO/IEC 23053 AI ML Framework. Japan AI Guidelines Safety + AISI applies.
The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.