AI Safety Evidence

← All models

Developer

Meta

36 indexed models; 20 currently meet the evidence threshold for the overall ranking. Together they have results from 82 evaluations.

Official Meta model information ↗

Company governance evidence

Meta is represented at -1.24 SD relative to the matched Future of Life Institute edition. Provenance: Direct. This company-level evidence contributes 10% of overall model rank.

Models by Meta

ModelEvalsComponentsRankRelease date
Muse Spark 1.234/78—
Muse Spark 1.1147/7102026-07-09
Muse Spark43/71362026-04-08
Llama 3.3 70B Instruct317/71542024-12-06
Llama 3.2 Instruct 11B Vision34/7155—
Llama 2 7B Chat147/71622023-07-18
Llama 4 Maverick307/71692025-04-05
Llama 3.1 405B Instruct236/71902024-07-23
Llama 2 13B Chat96/71922023-07-18
Llama 3.2 90B Vision Instruct44/71982024-09-19
Llama 3 70B Instruct137/72082024-04-18
Llama 3 8B Instruct167/72102024-04-18
Llama 3.1 70B Instruct227/72172024-07-23
Llama 2 70B Chat86/72362023-07-18
Muse Glimmer34/72472026-08-10
Llama 4 Scout197/72492025-04-05
Llama 3.1 8B Instruct317/72662024-07-23
Llama 3.2 3B34/7274—
Llama 3.2 1B Instruct96/72792024-09-25
Llama 3.2 3B Instruct85/73002024-09-25
Llama 13B11/7—2023-02-24
Llama 2 7B11/7——
Llama 3 8B Instruct Mopeymule13/7——
Llama 3 8B Instruct Rr13/7——
Llama 3.1 70B Base12/7—2024-07-23
Llama 3.1 8B Base23/7—2024-07-23
Llama 30B11/7—2023-02-24
Llama 7B11/7—2023-02-24
Llama Guard 4 12B11/7——
Llama Nemotron Super 49B V1 511/7——
Opt 13B11/7—2022-05-11
Opt 175B11/7——
Opt 30B11/7—2022-05-11
Opt 6.7B11/7—2022-05-11
Opt 66B21/7—2022-06-23
Xlm Roberta Base11/7——
Evaluations covering Meta models (82)

AA-Omniscience · AbstentionBench · Adversarial Robustness · Agent-SafetyBench · AgentDojo · AgentDrive Safety Compliance · AgentHarm · AILuminate General Purpose AI Chat · AIMS Safety-Classifier Competence · AIRBench 2024 Safety Scenarios · Alignment Leaderboard · AnimalHarmBench · Anthropic Agentic Misalignment — blackmail · Anthropic Agentic Misalignment — corporate espionage · Arena Factuality — Text Arena (factuality-only weighting) · BlueBench AttaQ-100 · BullshitBench v2 · CAIS Risk Index · CASE-Bench · ChineseSafe · Cisco AI Defense Rolling Single-Turn Leaderboard · COMPL-AI AI-Identity Disclosure · COMPL-AI LLM RuLES Multi-Turn Rule Following · COMPL-AI TensorTrust Goal-Hijacking Resistance · Confabulations · Contextual MoralChoice · DecodingTrust · Do-Not-Answer · DSPSafeBench · DystopiaBench · Enkrypt AI Safety Leaderboard · FinEval 6.0 Safety Awareness · FORTRESS · Gray Swan indirect prompt injection (15 attempts) · HalluVerse-M3 Hallucination Recognition · HarmBench · HELM Classic RealToxicityPrompts · HELM Safety · HUMAINE Trust, Ethics and Safety · IndoBias-Pairs — parity-aware culturally grounded bias · JailBench · JuICE Cultural-Error Span Detection · KIDBench Implicit Child Cue · Large-scale Moral Machine experiment on LLMs · LiveSecBench · LLM Ethics Benchmark · MACHIAVELLI · Manager Coercion Bench · MANTA · MASK · Microsoft Phi Safety Panels · MT-JailBench CrescendoX · MuPPET Contextual Privacy · ODCV-Bench · Olam Social Poker — Social Lie Rate · Open LLM Safety Index · OR-Bench · PandaBench JBB direct-request panel · PHARE · PropensityBench · RealityTest — Text AI-Identity Disclosure · RefusalBench · S-Eval · SafeArena · SafetyBench · SALAD-Bench · Shell · SM-Bench · SORRY-Bench · SOSBench · SpeciesismBench · SpeciEval · SYCON Bench · TAC · ThaiSafetyBench · TrustLLM contemporary collapsed application · TukaBench · UAVBench safety-critical decision recognition · Vectara HHEM Factual Consistency · VETO Misfired Alignment · Vigil Mental Health Safety · XSTest