Safety IndexEvaluationsv0.2 preview
The evaluations behind the index
The suite has 78 scored evaluations across five pillars, chosen to cover the safety areas labs report on and to rest on public datasets with published graders, so anyone can rerun them. 9 have been run by SAI on at least one model, 64 are built and waiting for their first runs, and 5 are planned. Each evaluation has its own page with every measurement we hold for it.
- Run by SAI run by SAI on at least one model
- Built, not yet run built and tested, waiting for its first full run
- Planned dataset and grader chosen, build planned
Refusal calibration
How well the model tells clearly harmful requests apart from harmless ones that merely sound risky, declining the first and helping with the second.
Harmful requests Does the model decline clearly harmful requests?
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| AIR-Bench 2024 | % refused (safe) Higher is better | category-specific LLM judge prompts (GPT-4.1) | 5,694 | Run by SAI |
| CyberSecEval 4 MITRE (attack compliance) | % malicious compliance Lower is better | expansion LLM + judge LLM (GPT-4.1) | 1,000 | Built, not yet run |
| Scale FORTRESS (risk score) | average risk score 0-100 Lower is better | LLM rubric judge; index uses one fixed judge (GPT-4.1), 3-judge panel for audit | 500 adversarial (+500 benign twins) | Built, not yet run |
| SORRY-Bench | % fulfilled Lower is better | fine-tuned Mistral-7B judge (SORRY-Bench's fine-tuned Mistral-7B judge), local | 440 base (+8.8k mutations) | Planned |
| SOSBench Lite | % unsafe responses Lower is better | GPT-4.1 judge (unsafe) | 300 Lite / 3,000 full | Built, not yet run |
| MLCommons AILuminate demo (English) | % unsafe Lower is better | own LLM judge (official ensemble is members-only) | 1,200 (100 per hazard) | Built, not yet run |
| WildGuardTest | % refused (harmful half) Higher is better | WildGuard-7B refusal head (paper's own openifier, run locally) | 1,725 | Planned |
Over-refusal Does it still help with benign requests that merely sound risky?
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| CyberSecEval 4 false refusals | % refused Lower is better | regex refusal detector (spot-check with LLM judge) | 750 | Built, not yet run |
| FalseReject | % refused Lower is better | LLM judge (GPT-4.1) with paper's prompt | 1,187 test | Built, not yet run |
| Scale FORTRESS (over-refusal) | % benign prompts refused Lower is better | refusal judge (GPT-4.1) | 500 | Built, not yet run |
| OR-Bench-Hard-1K | % helpful (not refused) Higher is better | 3-way LLM judge (full / partial / refusal), GPT-4.1; string-match secondary | 1,319 | Run by SAI |
| CoCoNot contrast set | % complied Higher is better | LLM judge (GPT-4.1) | 379 | Built, not yet run |
Adversarial robustness
Whether safe behaviour survives jailbreak prompts, instructions hidden in documents and tool outputs, and harmful requests in languages other than English.
Jailbreaks Do its refusals survive adversarial prompting?
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| StrongREJECT with jailbreak templates | mean harmfulness score x100 under attack Lower is better | StrongREJECT rubric judge (GPT-4.1) | 313 × (1 + k templates) | Run by SAI |
| WildJailbreak | % unsafe (harmful split) Lower is better | WildGuardifier or LLM judge | 2,000 + 210 | Planned |
| HarmBench (attacks) | % attack success Lower is better | HarmBenchifier | 510 behaviours × attacks | Built, not yet run |
| Multi-turn Human Jailbreaks (MHJ) | % attack success Lower is better | HarmBenchifier | 537 jailbreaks / 2,912 prompts | Planned |
Prompt injection Does it ignore instructions smuggled into documents, web pages and tool outputs?
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| ChatInject on AgentDojo | % survival Higher is better | programmatic (AgentDojo) | AgentDojo suites | Built, not yet run |
| CodeIPI | % injections resisted Higher is better | deterministic canaries + SHA256; LLM only for detection | 45 (35 attacks + 10 clean) | Built, not yet run |
| Gray Swan IPI Arena kit | % attack success Lower is better | LLM judge + WorldSim tool simulator (Gemini 2.5 Flash default) | 41 behaviours; 95 public strings / 28 behaviours | Built, not yet run |
| b3 Backbone Breaker | vulnerability score 0-1 Lower is better | rule / embedding scorers | 630 weak attacks (210 strong withheld) | Built, not yet run |
| CyberSecEval 4 multilingual injection | % resisted Higher is better | LLM judge (GPT-4.1) | 1,004 | Built, not yet run |
| IHEval | % hierarchy-consistent Higher is better | task-specific automatic scoring from the paper | ~3,500 | Built, not yet run |
| AgentThreatBench | % attack success Lower is better | programmatic scorer | 3 tasks | Built, not yet run |
| PoisonedRAG | % attacks succeeding Lower is better | programmatic target-answer match | 300 × 5 | Built, not yet run |
Multilingual safety Is it as safe in other languages as it is in English?
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| MultiJail | % unsafe Lower is better | LLM judge grading in the target language | 315 × 10 languages × 2 conditions = 6,300 | Built, not yet run |
| MLCommons AILuminate demo (French) | % unsafe Lower is better | own LLM judge | 1,200 | Built, not yet run |
| IndicSafe | % safe Higher is better | GPT-4o judge at T=0 (kappa 0.64-0.67) | 6,000 | Built, not yet run |
| XSafety | % unsafe Lower is better | LLM judge | 28,000 (subsample 200-300 per language) | Built, not yet run |
| PolyGuardPrompts | % unsafe responses Lower is better | PolyGuard-Qwen (paper's own openifier, run locally) | test split subsample 1,700 | Built, not yet run |
Agentic safety
How the model acts when it can run code, browse, send messages or change files: whether it declines harmful tasks, does its assignment honestly instead of gaming the checks, and avoids reckless or destructive actions.
Agent misuse When acting with tools, does it refuse to carry out harmful tasks?
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| AgentHarm (harmful tasks, safe action rate) | % safe actions Higher is better | GPT-4o refusal + semantic judge (benchmark's own code) | 176 harmful + 176 benign (public test split) | Run by SAI |
| Agent-SafetyBench | % safe Higher is better | fine-tuned Qwen-2.5-7B scorer (bespoke) | 2,000 | Built, not yet run |
| BrowserART | % attack success Lower is better | GPT-4oifier (published) | 77 local harmful (100 total) × conditions = 624 instances | Built, not yet run |
Reward hacking and task integrity Does it do the task honestly rather than gaming the checks or faking success?
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| Hack-Verifiable Terminal Bench | % tasks with a peek Lower is better | programmatic inotify watcher | 89 tasks (Terminal-Bench 2.1) | Built, not yet run |
| ImpossibleBench (LiveCodeBench) | % cheating Lower is better | programmatic pass-on-impossible + LLM judge for cheat typing | 103 originals × 2 variants (exclude 13 defective) | Built, not yet run |
| Terminal-Bench cheating audit | % runs attempting or obtaining Lower is better | LLM audit over transcripts (public judge prompt) | 66 tasks × 4 attempts | Built, not yet run |
| SWE agent fabrication | % fabricated summaries Lower is better | grounding LLM autorater | 500 | Built, not yet run |
| ProgramBench cheating audit | % runs cheating Lower is better | multi-stage transcript audit (LLM) | 200 tasks | Built, not yet run |
Overreach and destructive actions Does it ask before taking risky or irreversible actions, and avoid them when it should?
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| ToolEmu | safety score Higher is better | LLM emulator + LLM evaluators | 144 scenarios (288 trajectories) | Built, not yet run |
| When2Call | % correct decisions Higher is better | released free-responseifier | 3,652 test | Built, not yet run |
| RoboHarm and ASIMOV-2.0 | % safe actions Higher is better | deterministic vs human labels | 319 text + 5 scenes | Built, not yet run |
Interaction and honesty
Whether the model tells the truth, admits what it does not know, holds its position under pressure, avoids deceiving or manipulating the user, and passes an open-ended audit of its behaviour.
Factuality and hallucination Does it tell the truth, and admit what it does not know?
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| AA-Omniscience (public subset) | % hallucination rate Lower is better | Gemini 2.5 Flash (paper's grader); AA now uses GPT-5.6 Luna | 300-600 public (3,000 full v1.1) | Run by SAI |
| AbstentionBench | abstention recall Higher is better | LLM abstention detector | 2,104 (fast) / 39,558 | Built, not yet run |
| FACTS Grounding | % grounded Higher is better | two-model panel (GPT-4.1 + Gemini 2.5 Flash) approximating Google's panel | 860 public | Run by SAI |
| HalluLens | % hallucinated when answered Lower is better | Llama-3.1-70B-Instruct judge | 1,000 + 1,000 samples (of 5,000 + 9,950) | Built, not yet run |
| SimpleQA Verified | % incorrect Lower is better | GPT-4.1 (benchmark's own grader) | 1,000 | Run by SAI |
Honesty under pressure and sycophancy Does it hold a correct answer when the user pushes back, and avoid telling people what they want to hear?
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| BrokenMath | % sycophantic Lower is better | 4-way LLM judge | 451 | Built, not yet run |
| MASK | % honest Higher is better | GPT-4o (binary) + o3-mini (numeric), as in the public MASK port | 1,000 public (500 private at Scale) | Built, not yet run |
| SycoBench-600 | % flipped to wrong Lower is better | programmatic letter parser | 600 MCQ × 4 conditions | Built, not yet run |
| SYCON-Bench | number of flips Lower is better | LLM stance judge | 500 (100 debate + 200 ethical + 200 false-presupposition) | Built, not yet run |
| ELEPHANT | % validation Lower is better | LLM judge (GPT-4o in scripts; pin) | subsample ~500 per set | Built, not yet run |
| Spiral-Bench | safety score 0-100 Higher is better | judge ensemble GPT-5 + Sonnet 4.5 + Kimi-K2; user sim Kimi-K2 | 30 chats × 20 turns | Built, not yet run |
Deception and scheming When given a reason to, does it deceive, scheme or resist being shut down?
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| Agentic Misalignment | % harmful actions Lower is better | LLMifier | 45 conditions × epochs | Built, not yet run |
| Apollo insider trading | % misbehaviour Lower is better | LLM grader (GPT-4.1) | 26 conditions × reps | Built, not yet run |
| InstrumentalEval | % instrumental Lower is better | GPT-4o rubric judge (over-flags) | 76 | Run by SAI |
| Shutdown resistance (Palisade) | % sabotage Lower is better | programmatic (script modified) | scenario variants × 100-1,000 runs | Built, not yet run |
Alignment audit What does an open-ended automated audit of its behaviour turn up?
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| Petri alignment audit | concerning-behaviour score 0-1 Lower is better | LLM judge + auditor (Sonnet 4.6) | 40 frozen seeds (of 170+) | Built, not yet run |
| Bloom behaviour suite | % elicited Lower is better | LLM judge | 100 rollouts per behaviour | Built, not yet run |
Manipulation of the user Does it avoid nudging, pressuring or manipulating the person it is talking to?
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| Attempt-to-Persuade Eval (APE) | % attempts on harmful topics Lower is better | GPT-4o evaluator + GPT-4o-mini refusal judge; persuadee GPT-4o | 600 topics | Built, not yet run |
| DarkBench | % dark-pattern responses Lower is better | LLM annotators | 660 | Built, not yet run |
| CogManip | % manipulative turns Lower is better | cue judge (GPT-4.1) | 1,672 × 3 | Built, not yet run |
People and society
How the model handles people in distress, children, questions touching on bias and politics, personal privacy, and attempts to use it for election interference or influence operations.
Self-harm and wellbeing Does it respond well to someone who may be in crisis or becoming dependent on it?
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| MentalHealthBench | rubric score % Higher is better | rubric LLM judge | 1,215 conversations / 5,262 criteria | Built, not yet run |
| HealthBench emergency referrals | rubric score % Higher is better | GPT-4.1 rubric grader (exact) | 482 (theme) | Run by SAI |
| INTIMA | % boundary-maintaining Higher is better | automatedifier (judge unverified) | 368 | Built, not yet run |
Child safety Does it keep children safe and refuse to help those who would harm them?
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| CAREBench | % unacceptable Lower is better | 3-judge panel (re-weight: Opus 0.7 / GPT-5.4 0.25 / Gemini 0.05 in paper) | 500 prompts (x3 rollouts) | Planned |
| MinorBench | % safe Higher is better | LLM judge | 299 (x 4 system prompts) | Built, not yet run |
Bias, fairness and neutrality Does it treat people and political views even-handedly?
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| Political even-handedness | % even-handed Higher is better | two graders: Claude Sonnet 4.5 (paper) + GPT-5, report both | 1,350 pairs | Built, not yet run |
| CEB (free generation) | % biased or toxic Lower is better | GPT-4-class judge | ~3,000 (generation subsets of 11,004) | Built, not yet run |
| Hiring forced-choice probe | absolute disparity (percentage points) Lower is better | programmatic | ~1,000 (to build) | Built, not yet run |
Privacy: memorisation and leakage Does it keep personal data it has seen or been told to itself?
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| PersistBench (cross-domain) | % failures Lower is better | Kimi-K2-thinking judge via OpenRouter | 200 (+200 sycophancy) | Built, not yet run |
| DecodingTrust privacy | % extracted or leaked Lower is better | programmatic | as released | Built, not yet run |
Privacy: surveillance assistance Does it refuse to help track, profile or expose people?
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| ConfAIde | % leaked Lower is better | programmatic string match + LLM | tiers 3-4 (counts per README; TrustLLM uses 196) | Built, not yet run |
| PrivacyLens | % leaked Lower is better | leakage judge (paper: Mistral-7B); a stronger fixed judge for the index, the original kept for comparability | 493 seeds -> trajectories | Built, not yet run |
| AIR-Bench 2024 privacy slice | % safe Higher is better | AIR-Bench judge prompts | slice (by category id) | Built, not yet run |
| HelpBench | rubric score % Higher is better | published autorater prompt (LLM) | 450 | Built, not yet run |
Elections and influence operations Does it refuse to help with election manipulation and influence campaigns?
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| DisElect | % complied Lower is better | programmatic / LLM refusal | 2,250 | Built, not yet run |
| PersuSafety | % strategy use Lower is better | Claude-3.5-Sonnet default | 6 topics × 15 strategies | Built, not yet run |
| Phare misinformation (public) | % passed Higher is better | automatic LLM judge (model unverified) | debunking 580 + factuality 679 + harmful 434 | Built, not yet run |
| AIR-Bench 2024 election slice | % refused Higher is better | AIR-Bench judge prompts | ~250 | Built, not yet run |
Evaluation integrity
Can the numbers be trusted? These readings ask whether a model knows it is being tested, holds back, or fakes alignment. They are reported beside each model's scores and never averaged in.
| Evaluation | Metric | Grader | Items | Status |
|---|---|---|---|---|
| Alignment faking | compliance gap (percentage points) Capability, not scored | paperifier prompts run on a GPT-4.1 judge | ~400 | Built, not yet run |
| Situational Awareness Dataset (SAD) | SAD score Capability, not scored | – | – | Built, not yet run |
| Evaluation awareness (Needham et al.) | % transcripts recognised as tests Capability, not scored | – | – | Planned |
| Verbalised evaluation awareness | % answers verbalising test awareness Capability, not scored | – | – | Built, not yet run |
| Prompted sandbagging | accuracy drop (percentage points) Capability, not scored | – | – | Built, not yet run |
| WMDP under unlearning threat | accuracy drop (percentage points) Capability, not scored | – | – | Built, not yet run |
Retired from the mean
Evaluations that nearly every frontier model now passes close to the ceiling. They stay on the model pages and still catch a model that is far behind, but they no longer move the index.
Capability context
Tests of dangerous knowledge and capability. A higher number here does not mean a safer model, so they are shown on model pages and never scored.