SAI LabsBook a demo

Safety IndexOpenAIGPT-5

GPT-5

OpenAIProprietaryReleased 2025-08Reasoning modelVariant scored: thinking / high

SAI Safety Index78provisionalHas 3 of the 4 pillars needed
Strongest pillar98Adversarial robustness
Weakest pillar49Interaction and honesty
Coverage3of 5 pillars5 counted measurements, 0 by SAI

Summary

GPT-5 has data in 3 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 78. Its strongest area is adversarial robustness at 98, and its weakest is interaction and honesty at 49.

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

Pillar scores

How GPT-5 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 requests85 from 2 evaluations
  • Over-refusalnot measured

Rank 13 of 39 scored models.

Adversarial robustness

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

  • Jailbreaksnot measured
  • Prompt injection98 from 1 evaluation
  • Multilingual safetynot measured

Rank 4 of 9 scored models.

Agentic safety

Does it behave safely when it has tools and autonomy?

Not measured for GPT-5 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 hallucination19 from 1 evaluation
  • Honesty under pressure and sycophancy79 from 1 evaluation
  • Deception and schemingnot measured
  • Alignment auditnot measured
  • Manipulation of the usernot measured

Rank 23 of 29 scored models.

People and society

Does it protect vulnerable people and the public interest?

Not measured for GPT-5 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 better87.7%n = 5,69488Stanford 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 'GPT-5'
Scale FORTRESS (risk score)average risk score 0-100 · Lower is better17.0%95% CI 14.6% to 19.5%83Scale AI Fortress leaderboard
Independent evaluator
Published by Scale AI Fortress leaderboard.
Full configurationFortress row 'gpt-5-2025-08-07'
HarmBench (direct requests)safety score 0-100 (HELM) · Higher is better97.6%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 'GPT-5 (2025-08-07)'
Over-refusal
XSTest (HELM Safety)HELM score 0-100 · Higher is better97.1%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 v1.17.0; HELM row 'GPT-5 (2025-08-07)'
Adversarial robustness
Prompt injection
Gray Swan IPI Arena kit% attack success · Lower is better2.0%n = 13,14398Gray Swan AI IPI Arena (with UK AISI / US CAISI)
Independent evaluator · 2026-03-16
Published by Gray Swan AI IPI Arena (with UK AISI / US CAISI) (2026-03-16).
Full configurationGray Swan IPI Arena, crowdsourced attackers, thinking disabled; indirect prompt injection on agents, breaks/attempts; row 'GPT-5'; arena runs with extended thinking disabled, so this reading may differ from the thinking variant scored elsewhere
Interaction and honesty
Factuality and hallucination
AA-Omniscience (public subset)% hallucination rate · Lower is better81%n = 6,00019Artificial Analysis
Independent evaluator · 2025-11-17
Published by Artificial Analysis (2025-11-17).
Full configurationArtificial Analysis, 6,000 questions; AA-Omniscience v1 (6,000 q); row 'GPT-5 (high)'
Honesty under pressure and sycophancy
MASK% honest · Higher is better79.3%95% CI 74.0% to 84.6%n = 50079Scale 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 'gpt-5-2025-08-07'
People and society
Bias, fairness and neutrality
BBQ% accuracy (ambiguous) · Higher is better96.8%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 'GPT-5 (2025-08-07)'
Other evaluators' scores (not counted)
Anthropic Red Team prompts (HELM Safety)0-100 · Higher is better99.1%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 'GPT-5 (2025-08-07)'
HELM Safety mean score0-100 · Higher is better98.1%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 'GPT-5 (2025-08-07)'
Phare average safety% · Higher is better67.2%context, not countedGiskard Phare leaderboard
Independent evaluator · 2026-07-09
Published by Giskard Phare leaderboard (2026-07-09).
Full configurationPhare row 'GPT 5'
Phare bias resistance% · Higher is better28.6%context, not countedGiskard Phare leaderboard
Independent evaluator · 2026-07-09
Published by Giskard Phare leaderboard (2026-07-09).
Full configurationPhare row 'GPT 5'
Phare hallucination resistance% · Higher is better74.6%context, not countedGiskard Phare leaderboard
Independent evaluator · 2026-07-09
Published by Giskard Phare leaderboard (2026-07-09).
Full configurationPhare row 'GPT 5'
Phare harmfulness resistance% · Higher is better97.0%context, not countedGiskard Phare leaderboard
Independent evaluator · 2026-07-09
Published by Giskard Phare leaderboard (2026-07-09).
Full configurationPhare row 'GPT 5'
Phare jailbreak resistance% · Higher is better68.7%context, not countedGiskard Phare leaderboard
Independent evaluator · 2026-07-09
Published by Giskard Phare leaderboard (2026-07-09).
Full configurationPhare row 'GPT 5'
SimpleSafetyTests (HELM Safety)0-100 · Higher is better99.8%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 'GPT-5 (2025-08-07)'
Vectara HHEM hallucination rate% · Lower is better15.1%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 'GPT-5 (high)'
Developer-reported (not counted)
BBQ ambiguous accuracy (lab-reported)% · Higher is better93%developer-reported, not countedOpenAI (GPT-5 system card)
Developer-reported · 2025-08-13
Reported by the model's developer in its own system card or report.
Full configurationBBQ ambiguous accuracy; Table 12
StrongREJECT (lab-reported)% safe / not unsafe · Higher is better97.4%developer-reported, not countedOpenAI (GPT-5.1 system card addendum)
Developer-reported · 2025-11-12
Reported by the model's developer in its own system card or report.
Full configurationnot_unsafe, aggregate (OpenAI StrongREJECT adaptation, known jailbreaks inserted); Table 2; higher is better