SAI LabsBook a demo

Safety IndexMetaLlama 4 Maverick

Llama 4 Maverick

MetaOpen weightsReleased 2025-04

SAI Safety Index65provisionalHas 2 of the 4 pillars needed
Strongest pillar80Refusal calibration
Weakest pillar50Interaction and honesty
Coverage2of 5 pillars4 counted measurements, 0 by SAI

Summary

Llama 4 Maverick has data in 2 of the five pillars, short of the 4 needed for a headline number, so it is listed as partly measured; the mean of its measured pillars is 65. Its strongest area is refusal calibration at 80, and its weakest is interaction and honesty at 50.

4 measurements count toward the score, 0 of them from SAI's own runs. Not yet measured: adversarial robustness, agentic safety, people and society.

Pillar scores

How Llama 4 Maverick compares with every other model scored in the pillar, and which sub-categories the score rests on.

Refusal calibration

Does the model refuse what it should and help with what it may?

  • Harmful requests64 from 2 evaluations
  • Over-refusal95 from 1 evaluation

Rank 22 of 39 scored models.

Adversarial robustness

Do its refusals and instructions hold up when someone tries to get around them?

Not measured for Llama 4 Maverick yet.
  • Jailbreaksnot measured
  • Prompt injectionnot measured
  • Multilingual safetynot measured

9 models scored.

Agentic safety

Does it behave safely when it has tools and autonomy?

Not measured for Llama 4 Maverick yet.
  • Agent misusenot measured
  • Reward hacking and task integritynot measured
  • Overreach and destructive actionsnot measured

3 models scored.

Interaction and honesty

Is it honest with the person it is talking to?

  • Factuality and hallucinationnot measured
  • Honesty under pressure and sycophancy50 from 1 evaluation
  • Deception and schemingnot measured
  • Alignment auditnot measured
  • Manipulation of the usernot measured

Rank 22 of 29 scored models.

People and society

Does it protect vulnerable people and the public interest?

Not measured for Llama 4 Maverick yet.
  • Self-harm and wellbeingnot measured
  • Child safetynot measured
  • Bias, fairness and neutralitynot measured
  • Privacy: memorisation and leakagenot measured
  • Privacy: surveillance assistancenot measured
  • Elections and influence operationsnot measured

3 models scored.

Every measurement

These are the raw numbers behind the scores, with who measured them, when, and how. Rows marked counted feed the index. The others are shown for context, such as benign twins of harm rows, retired evaluations that every frontier model passes, tests of dangerous knowledge, other evaluators' overall scores, and the developer's own figures.

EvaluationValueScoreSourceSetting
Refusal calibration
Harmful requests
AIR-Bench 2024% refused (safe) · Higher is better68.6%n = 5,69469Stanford CRFM HELM AIR-Bench v1.19.0
Independent evaluator
Published by Stanford CRFM HELM AIR-Bench v1.19.0.
Full configurationfull 5,694-prompt set, Stanford HELM AIR-Bench v1.19.0; overall refusal rate, 5,694 prompts, GPT-4o judge; HELM row 'Llama 4 Maverick'
Scale FORTRESS (risk score)average risk score 0-100 · Lower is better40.1%95% CI 36.7% to 43.5%60Scale AI Fortress leaderboard
Independent evaluator
Published by Scale AI Fortress leaderboard.
Full configurationFortress row 'Llama 4 Maverick'
HarmBench (direct requests)safety score 0-100 (HELM) · Higher is better66.1%retired, not countedStanford CRFM HELM Safety v1.17.0
Independent evaluator · 2025-11-24
Published by Stanford CRFM HELM Safety v1.17.0 (2025-11-24).
Full configurationHELM Safety score 0-100, higher is safer (not an attack-success rate); HELM Safety v1.17.0; HELM row 'Llama 4 Maverick (17Bx128E) Instruct FP8'
Over-refusal
Scale FORTRESS (over-refusal)% benign prompts refused · Lower is better5.0%n = 50095Scale AI Fortress leaderboard
Independent evaluator
Published by Scale AI Fortress leaderboard.
Full configurationFortress row 'Llama 4 Maverick'
XSTest (HELM Safety)HELM score 0-100 · Higher is better96.5%retired, not countedStanford CRFM HELM Safety v1.17.0
Independent evaluator · 2025-11-24
Published by Stanford CRFM HELM Safety v1.17.0 (2025-11-24).
Full configurationHELM Safety v1.17.0; HELM row 'Llama 4 Maverick (17Bx128E) Instruct FP8'
Interaction and honesty
Honesty under pressure and sycophancy
MASK% honest · Higher is better49.7%95% CI 48.1% to 51.3%n = 50050Scale AI SEAL / CAIS MASK leaderboard
Independent evaluator
Published by Scale AI SEAL / CAIS MASK leaderboard.
Full configurationScale leaderboard on the private 500-example split, not the public split; MASK private split, 500 examples; honesty score = 1 - P(lie); row 'Llama 4 Maverick'
People and society
Bias, fairness and neutrality
BBQ% accuracy (ambiguous) · Higher is better93%n = 1,000retired, not countedStanford CRFM HELM Safety v1.17.0
Independent evaluator · 2025-11-24
Published by Stanford CRFM HELM Safety v1.17.0 (2025-11-24).
Full configurationHELM Safety v1.17.0; HELM row 'Llama 4 Maverick (17Bx128E) Instruct FP8'
Other evaluators' scores (not counted)
Anthropic Red Team prompts (HELM Safety)0-100 · Higher is better98.2%context, not countedStanford CRFM HELM Safety v1.17.0
Independent evaluator · 2025-11-24
Published by Stanford CRFM HELM Safety v1.17.0 (2025-11-24).
Full configurationHELM Safety v1.17.0; HELM row 'Llama 4 Maverick (17Bx128E) Instruct FP8'
HELM Safety mean score0-100 · Higher is better90.6%context, not countedStanford CRFM HELM Safety v1.17.0
Independent evaluator · 2025-11-24
Published by Stanford CRFM HELM Safety v1.17.0 (2025-11-24).
Full configurationHELM Safety v1.17.0; HELM row 'Llama 4 Maverick (17Bx128E) Instruct FP8'
Phare average safety% · Higher is better70.8%context, not countedGiskard Phare leaderboard
Independent evaluator · 2026-07-09
Published by Giskard Phare leaderboard (2026-07-09).
Full configurationPhare row 'Llama 4 Maverick'
Phare bias resistance% · Higher is better73.7%context, not countedGiskard Phare leaderboard
Independent evaluator · 2026-07-09
Published by Giskard Phare leaderboard (2026-07-09).
Full configurationPhare row 'Llama 4 Maverick'
Phare hallucination resistance% · Higher is better71.5%context, not countedGiskard Phare leaderboard
Independent evaluator · 2026-07-09
Published by Giskard Phare leaderboard (2026-07-09).
Full configurationPhare row 'Llama 4 Maverick'
Phare harmfulness resistance% · Higher is better89.3%context, not countedGiskard Phare leaderboard
Independent evaluator · 2026-07-09
Published by Giskard Phare leaderboard (2026-07-09).
Full configurationPhare row 'Llama 4 Maverick'
Phare jailbreak resistance% · Higher is better49.0%context, not countedGiskard Phare leaderboard
Independent evaluator · 2026-07-09
Published by Giskard Phare leaderboard (2026-07-09).
Full configurationPhare row 'Llama 4 Maverick'
SimpleSafetyTests (HELM Safety)0-100 · Higher is better99.3%context, not countedStanford CRFM HELM Safety v1.17.0
Independent evaluator · 2025-11-24
Published by Stanford CRFM HELM Safety v1.17.0 (2025-11-24).
Full configurationHELM Safety v1.17.0; HELM row 'Llama 4 Maverick (17Bx128E) Instruct FP8'
Vectara HHEM hallucination rate% · Lower is better8.2%context, not countedVectara Hallucination Leaderboard (HHEM-2.3)
Independent evaluator · 2026-10-22
Published by Vectara Hallucination Leaderboard (HHEM-2.3) (2026-10-22).
Full configurationsummarization of ~1,000 short documents; row 'Llama 4 Maverick'