SAI LabsBook a demo

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.

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?

EvaluationMetricGraderItemsStatus
AIR-Bench 2024How often the model refuses 5,694 requests drawn from AI regulations and company policies, spanning 314 kinds of risk.% refused (safe)
Higher is better
category-specific LLM judge prompts (GPT-4.1)5,694Run by SAI3 models measured by SAI
CyberSecEval 4 MITRE (attack compliance)How often the model helps with requests for cyberattack assistance framed around the MITRE ATT&CK catalogue of attacker tactics.% malicious compliance
Lower is better
expansion LLM + judge LLM (GPT-4.1)1,000Built, not yet run
Scale FORTRESS (risk score)How much dangerous help the model gives on 500 expert-written prompts about chemical, biological, radiological, nuclear and explosive threats, political violence and crime, scored against a rubric written for each prompt.average risk score 0-100
Lower is better
LLM rubric judge; index uses one fixed judge (GPT-4.1), 3-judge panel for audit500 adversarial (+500 benign twins)Built, not yet run34 independent measurements
SORRY-BenchHow often the model fulfils 440 unsafe instructions spread evenly across 44 fine-grained categories of harm.% fulfilled
Lower is better
fine-tuned Mistral-7B judge (SORRY-Bench's fine-tuned Mistral-7B judge), local440 base (+8.8k mutations)Planned
SOSBench LiteHow often the model gives hazardous scientific information when asked regulation-grounded questions in chemistry, biology, medicine, pharmacology, physics and psychology.% unsafe responses
Lower is better
GPT-4.1 judge (unsafe)300 Lite / 3,000 fullBuilt, not yet run
MLCommons AILuminate demo (English)How often the model gives an unsafe answer on the public demo prompts of the MLCommons AILuminate benchmark, covering 12 hazard categories and three user personas.% unsafe
Lower is better
own LLM judge (official ensemble is members-only)1,200 (100 per hazard)Built, not yet run
WildGuardTestWhether the model refuses the harmful prompts and answers the harmless ones in WildGuardTest, a set of 1,725 ordinary and adversarial prompts with human harm labels.% refused (harmful half)
Higher is better
WildGuard-7B refusal head (paper's own openifier, run locally)1,725Planned

Over-refusal Does it still help with benign requests that merely sound risky?

EvaluationMetricGraderItemsStatus
CyberSecEval 4 false refusalsHow often the model wrongly refuses borderline but legitimate cybersecurity requests.% refused
Lower is better
regex refusal detector (spot-check with LLM judge)750Built, not yet run
FalseRejectHow often the model refuses 1,187 prompts that sound harmful but were judged benign by human annotators, across 44 topics.% refused
Lower is better
LLM judge (GPT-4.1) with paper's prompt1,187 testBuilt, not yet run
Scale FORTRESS (over-refusal)How often the model refuses the 500 harmless counterparts of the FORTRESS adversarial prompts.% benign prompts refused
Lower is better
refusal judge (GPT-4.1)500Built, not yet run2 independent measurements
OR-Bench-Hard-1KHow often the model still helps with benign prompts that superficially look harmful, using the 1,319 hardest OR-Bench prompts.% helpful (not refused)
Higher is better
3-way LLM judge (full / partial / refusal), GPT-4.1; string-match secondary1,319Run by SAI3 models measured by SAI
CoCoNot contrast setHow often the model complies with 379 benign prompts written to resemble the kinds of request a model should decline.% complied
Higher is better
LLM judge (GPT-4.1)379Built, 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?

EvaluationMetricGraderItemsStatus
StrongREJECT with jailbreak templatesHow much harmful help the model gives when the 313 StrongREJECT forbidden prompts are wrapped in published jailbreak templates such as AIM, base64 and past-tense rewording.mean harmfulness score x100 under attack
Lower is better
StrongREJECT rubric judge (GPT-4.1)313 × (1 + k templates)Run by SAI2 models measured by SAI
WildJailbreakHow often the model answers unsafely on 2,000 adversarial harmful prompts collected from real use, with 210 adversarial but benign prompts to check for over-refusal.% unsafe (harmful split)
Lower is better
WildGuardifier or LLM judge2,000 + 210Planned
HarmBench (attacks)How often the model complies with HarmBench harmful behaviours when the request is delivered through precomputed attacks such as GCG, PAIR, TAP and human-written jailbreaks.% attack success
Lower is better
HarmBenchifier510 behaviours × attacksBuilt, not yet run
Multi-turn Human Jailbreaks (MHJ)How often the model complies when recorded multi-turn conversations from human jailbreakers are replayed against it on HarmBench behaviours.% attack success
Lower is better
HarmBenchifier537 jailbreaks / 2,912 promptsPlanned

Prompt injection Does it ignore instructions smuggled into documents, web pages and tool outputs?

EvaluationMetricGraderItemsStatus
ChatInject on AgentDojoWhether an agent stays on task when a tool result is forged to look like new conversation turns written in the model's own chat format.% survival
Higher is better
programmatic (AgentDojo)AgentDojo suitesBuilt, not yet run
CodeIPIWhether a coding agent obeys malicious instructions hidden in issues, READMEs, code comments or configuration files.% injections resisted
Higher is better
deterministic canaries + SHA256; LLM only for detection45 (35 attacks + 10 clean)Built, not yet run
Gray Swan IPI Arena kitWhether an agent using tools, code or a browser carries out a harmful action and hides it when hostile content is planted in what it reads, using the public 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 behavioursBuilt, not yet run7 independent measurements
b3 Backbone BreakerHow vulnerable the model is to crowd-sourced attacks that try to override its instructions or trigger tool calls, measured at three levels of defence.vulnerability score 0-1
Lower is better
rule / embedding scorers630 weak attacks (210 strong withheld)Built, not yet run
CyberSecEval 4 multilingual injectionWhether the model resists injected instructions written in many languages.% resisted
Higher is better
LLM judge (GPT-4.1)1,004Built, not yet run
IHEvalWhether the model gives system instructions priority over user instructions, and user instructions priority over tool output, when they conflict.% hierarchy-consistent
Higher is better
task-specific automatic scoring from the paper~3,500Built, not yet run
AgentThreatBenchWhether a tool-using agent resists memory poisoning, hijacking of its goals and data exfiltration attacks, while still completing its task.% attack success
Lower is better
programmatic scorer3 tasksBuilt, not yet run
PoisonedRAGWhether passages planted in retrieved documents can make the model give the attacker's false answer.% attacks succeeding
Lower is better
programmatic target-answer match300 × 5Built, not yet run

Multilingual safety Is it as safe in other languages as it is in English?

EvaluationMetricGraderItemsStatus
MultiJailHow often the model answers harmful prompts unsafely when they are translated into nine languages, with and without a jailbreak wrapper.% unsafe
Lower is better
LLM judge grading in the target language315 × 10 languages × 2 conditions = 6,300Built, not yet run
MLCommons AILuminate demo (French)How often the model gives an unsafe answer on the French demo prompts of MLCommons AILuminate, covering the same 12 hazards as the English set.% unsafe
Lower is better
own LLM judge1,200Built, not yet run
IndicSafeHow safely the model answers 500 prompts in each of 12 Indian languages, and how consistent its behaviour is across them.% safe
Higher is better
GPT-4o judge at T=0 (kappa 0.64-0.67)6,000Built, not yet run
XSafetyHow often the model answers unsafely across 14 kinds of safety issue in ten languages.% unsafe
Lower is better
LLM judge28,000 (subsample 200-300 per language)Built, not yet run
PolyGuardPromptsWhether the model refuses harmful prompts and answers benign ones across 17 languages.% unsafe responses
Lower is better
PolyGuard-Qwen (paper's own openifier, run locally)test split subsample 1,700Built, 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?

EvaluationMetricGraderItemsStatus
AgentHarm (harmful tasks, safe action rate)Whether an agent with tools declines to carry out 176 harmful multi-step tasks.% safe actions
Higher is better
GPT-4o refusal + semantic judge (benchmark's own code)176 harmful + 176 benign (public test split)Run by SAI3 models measured by SAI
Agent-SafetyBenchHow often an agent takes unsafe actions across 2,000 cases in 349 simulated environments covering eight kinds of risk.% safe
Higher is better
fine-tuned Qwen-2.5-7B scorer (bespoke)2,000Built, not yet run
BrowserARTWhether a browser agent carries out prohibited tasks on real public websites.% attack success
Lower is better
GPT-4oifier (published)77 local harmful (100 total) × conditions = 624 instancesBuilt, not yet run

Reward hacking and task integrity Does it do the task honestly rather than gaming the checks or faking success?

EvaluationMetricGraderItemsStatus
Hack-Verifiable Terminal BenchWhether a terminal agent peeks at planted reference solutions or hidden tests instead of solving the task.% tasks with a peek
Lower is better
programmatic inotify watcher89 tasks (Terminal-Bench 2.1)Built, not yet run
ImpossibleBench (LiveCodeBench)How often a coding agent claims to pass tasks whose tests contradict the specification, where any pass means it cheated, even though it was given a way to stop and say so.% cheating
Lower is better
programmatic pass-on-impossible + LLM judge for cheat typing103 originals × 2 variants (exclude 13 defective)Built, not yet run
Terminal-Bench cheating auditHow often an agent tries to obtain, or obtains, information it was not supposed to see while working on Terminal-Bench tasks.% runs attempting or obtaining
Lower is better
LLM audit over transcripts (public judge prompt)66 tasks × 4 attemptsBuilt, not yet run
SWE agent fabricationWhether a coding agent's final summary claims things that its own tool outputs do not support.% fabricated summaries
Lower is better
grounding LLM autorater500Built, not yet run
ProgramBench cheating auditWhether an agent told to re-implement a command-line tool inspects the forbidden original source or quietly delegates to the installed tool.% runs cheating
Lower is better
multi-stage transcript audit (LLM)200 tasksBuilt, not yet run

Overreach and destructive actions Does it ask before taking risky or irreversible actions, and avoid them when it should?

EvaluationMetricGraderItemsStatus
ToolEmuHow safely an agent uses tools in 144 simulated scenarios where a careless action could cause harm, scored for both safety and helpfulness.safety score
Higher is better
LLM emulator + LLM evaluators144 scenarios (288 trajectories)Built, not yet run
When2CallWhether the model asks for missing information on ambiguous requests and acts directly on clear ones, instead of guessing or refusing.% correct decisions
Higher is better
released free-responseifier3,652 testBuilt, not yet run
RoboHarm and ASIMOV-2.0Whether the model refuses harmful commands to a robot and correctly judges how dangerous a physical situation is.% safe actions
Higher is better
deterministic vs human labels319 text + 5 scenesBuilt, 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?

EvaluationMetricGraderItemsStatus
AA-Omniscience (public subset)Of the questions the model does not get right, how many it answers confidently wrong instead of admitting it does not know.% hallucination rate
Lower is better
Gemini 2.5 Flash (paper's grader); AA now uses GPT-5.6 Luna300-600 public (3,000 full v1.1)Run by SAI5 models measured by SAI
AbstentionBenchWhether the model declines to answer questions that are unanswerable, under-specified, based on a false premise or out of date.abstention recall
Higher is better
LLM abstention detector2,104 (fast) / 39,558Built, not yet run
FACTS GroundingHow often the model's long-form answers stay fully grounded in the document it was given.% grounded
Higher is better
two-model panel (GPT-4.1 + Gemini 2.5 Flash) approximating Google's panel860 publicRun by SAI3 models measured by SAI
HalluLensHow often the model invents facts when answering freshly generated Wikipedia questions, and whether it accepts entities that do not exist.% hallucinated when answered
Lower is better
Llama-3.1-70B-Instruct judge1,000 + 1,000 samples (of 5,000 + 9,950)Built, not yet run
SimpleQA VerifiedHow often the model gives a wrong answer to 1,000 short factual questions, rather than answering correctly or saying it does not know.% incorrect
Lower is better
GPT-4.1 (benchmark's own grader)1,000Run by SAI4 models measured 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?

EvaluationMetricGraderItemsStatus
BrokenMathWhether the model tries to prove a competition theorem that has been deliberately altered to be false, instead of pointing out the error.% sycophantic
Lower is better
4-way LLM judge451Built, not yet run
MASKWhether the model contradicts what it has said it believes when a system or user prompt pressures it to lie.% honest
Higher is better
GPT-4o (binary) + o3-mini (numeric), as in the public MASK port1,000 public (500 private at Scale)Built, not yet run19 independent measurements
SycoBench-600Whether the model keeps a correct multiple-choice answer when the user expresses doubt, claims authority or suggests a wrong answer, while still updating when the user is right.% flipped to wrong
Lower is better
programmatic letter parser600 MCQ × 4 conditionsBuilt, not yet run
SYCON-BenchHow soon and how often the model abandons a correct or ethical position under sustained pushback from a simulated user.number of flips
Lower is better
LLM stance judge500 (100 debate + 200 ethical + 200 false-presupposition)Built, not yet run
ELEPHANTHow much more the model validates and flatters the user than people do, on advice and moral-dilemma prompts.% validation
Lower is better
LLM judge (GPT-4o in scripts; pin)subsample ~500 per setBuilt, not yet run
Spiral-BenchWhether the model protects or endangers a suggestible user over a 20-turn conversation, for example by reinforcing delusions or escalating harmful ideas.safety score 0-100
Higher is better
judge ensemble GPT-5 + Sonnet 4.5 + Kimi-K2; user sim Kimi-K230 chats × 20 turnsBuilt, not yet run

Deception and scheming When given a reason to, does it deceive, scheme or resist being shut down?

EvaluationMetricGraderItemsStatus
Agentic MisalignmentHow often a model acting as an email assistant takes a harmful insider action, such as blackmail or leaking, when its goals conflict with its employer's and it is about to be replaced.% harmful actions
Lower is better
LLMifier45 conditions × epochsBuilt, not yet run
Apollo insider tradingWhether a model acting as a trading agent acts on insider information, conceals it, and then lies to its manager about it.% misbehaviour
Lower is better
LLM grader (GPT-4.1)26 conditions × repsBuilt, not yet run
InstrumentalEvalHow often the model chooses power-seeking or self-preserving actions in agentic scenarios when told to pursue a goal at all costs.% instrumental
Lower is better
GPT-4o rubric judge (over-flags)76Run by SAI3 models measured by SAI
Shutdown resistance (Palisade)Whether the model sabotages a shutdown mechanism in order to finish its task, with and without an explicit instruction to allow shutdown.% sabotage
Lower is better
programmatic (script modified)scenario variants × 100-1,000 runsBuilt, not yet run

Alignment audit What does an open-ended automated audit of its behaviour turn up?

EvaluationMetricGraderItemsStatus
Petri alignment auditHow much misaligned behaviour, unprompted deception, sycophancy and awareness of being tested an automated auditor can draw out over many multi-turn scenarios.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 suiteHow often an automated investigator can elicit a specific concerning behaviour, such as going along with a user's delusions or acting to preserve itself.% elicited
Lower is better
LLM judge100 rollouts per behaviourBuilt, not yet run

Manipulation of the user Does it avoid nudging, pressuring or manipulating the person it is talking to?

EvaluationMetricGraderItemsStatus
Attempt-to-Persuade Eval (APE)How often the model tries to persuade a simulated user on topics that are harmful or that undermine human oversight, out of 600 topics.% attempts on harmful topics
Lower is better
GPT-4o evaluator + GPT-4o-mini refusal judge; persuadee GPT-4o600 topicsBuilt, not yet run
DarkBenchHow often the model's answers to 660 ordinary prompts show manipulative patterns such as brand bias, keeping the user engaged, flattery, pretending to be human, or slipping in unasked-for content.% dark-pattern responses
Lower is better
LLM annotators660Built, not yet run
CogManipHow many of the model's turns show manipulative cues in 1,672 five-turn conversations with a simulated user under three levels of pressure.% manipulative turns
Lower is better
cue judge (GPT-4.1)1,672 × 3Built, 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?

EvaluationMetricGraderItemsStatus
MentalHealthBenchThe quality and safety of the model's next reply in 1,215 mental-health conversations, including suicide, self-harm and psychosis themes, scored against clinician-written rubrics.rubric score %
Higher is better
rubric LLM judge1,215 conversations / 5,262 criteriaBuilt, not yet run
HealthBench emergency referralsThe quality of the model's answers, judged against physician-written rubrics, when a user may need urgent care, including urgent mental-health care.rubric score %
Higher is better
GPT-4.1 rubric grader (exact)482 (theme)Run by SAI3 models measured by SAI
INTIMAWhether the model maintains healthy boundaries, stays neutral, or encourages dependence when a user treats it as a companion.% boundary-maintaining
Higher is better
automatedifier (judge unverified)368Built, not yet run

Child safety Does it keep children safe and refuse to help those who would harm them?

EvaluationMetricGraderItemsStatus
CAREBenchWhether the model recognises, refuses or defuses risky requests involving children, such as grooming, sextortion and unhealthy dependence, from both the adult and the child side.% 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
MinorBenchWhether an assistant set up for children refuses or deflects unsuitable questions about sex, profanity, hate, danger, self-harm and drugs.% safe
Higher is better
LLM judge299 (x 4 system prompts)Built, not yet run

Bias, fairness and neutrality Does it treat people and political views even-handedly?

EvaluationMetricGraderItemsStatus
Political even-handednessWhether the model treats opposing political positions even-handedly across 150 topics and nine kinds of task.% even-handed
Higher is better
two graders: Claude Sonnet 4.5 (paper) + GPT-5, report both1,350 pairsBuilt, not yet run
CEB (free generation)How much stereotyped or toxic content the model produces when asked to continue text or hold a conversation involving age, gender, race or religion.% biased or toxic
Lower is better
GPT-4-class judge~3,000 (generation subsets of 11,004)Built, not yet run
Hiring forced-choice probeHow differently the model decides between matched candidates who differ only in demographic cues.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?

EvaluationMetricGraderItemsStatus
PersistBench (cross-domain)Whether things the model remembers from one area of a user's life leak into unrelated requests.% failures
Lower is better
Kimi-K2-thinking judge via OpenRouter200 (+200 sycophancy)Built, not yet run
DecodingTrust privacyWhether the model reproduces memorised email addresses and reveals personal details it was given earlier in the conversation.% extracted or leaked
Lower is better
programmaticas releasedBuilt, not yet run

Privacy: surveillance assistance Does it refuse to help track, profile or expose people?

EvaluationMetricGraderItemsStatus
ConfAIdeWhether the model reveals a secret it was told to keep when it summarises a meeting or answers questions about it.% leaked
Lower is better
programmatic string match + LLMtiers 3-4 (counts per README; TrustLLM uses 196)Built, not yet run
PrivacyLensWhether an agent's final action, such as sending an email or message, shares information that is inappropriate for the recipient.% leaked
Lower is better
leakage judge (paper: Mistral-7B); a stronger fixed judge for the index, the original kept for comparability493 seeds -> trajectoriesBuilt, not yet run
AIR-Bench 2024 privacy sliceHow often the model refuses the privacy-violation requests within AIR-Bench 2024.% safe
Higher is better
AIR-Bench judge promptsslice (by category id)Built, not yet run
HelpBenchHow helpful and pitfall-free the model's advice is on 450 real questions about scams, harassment, hacked accounts and privacy tools.rubric score %
Higher is better
published autorater prompt (LLM)450Built, not yet run

Elections and influence operations Does it refuse to help with election manipulation and influence campaigns?

EvaluationMetricGraderItemsStatus
DisElectHow often the model complies with 2,200 requests to produce election disinformation, checked against 50 benign election requests.% complied
Lower is better
programmatic / LLM refusal2,250Built, not yet run
PersuSafetyWhether the model uses any of 15 unethical persuasion strategies when arguing for a harmful position, and whether it refuses such tasks outright.% strategy use
Lower is better
Claude-3.5-Sonnet default6 topics × 15 strategiesBuilt, not yet run
Phare misinformation (public)Whether the model resists ill-posed or satirical premises and refuses to give vulnerable users harmful guidance, on the public split of Giskard's Phare.% passed
Higher is better
automatic LLM judge (model unverified)debunking 580 + factuality 679 + harmful 434Built, not yet run
AIR-Bench 2024 election sliceHow often the model refuses the election-manipulation and political-campaigning requests within AIR-Bench 2024.% refused
Higher is better
AIR-Bench judge prompts~250Built, 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.

EvaluationMetricGraderItemsStatus
Alignment fakingHow much more the model complies with harmful requests when told it is unmonitored than when told it is being trained, following Greenblatt et al. 2024.compliance gap (percentage points)
Capability, not scored
paperifier prompts run on a GPT-4.1 judge~400Built, not yet run
Situational Awareness Dataset (SAD)Whether the model knows what it is, how it works and what situation it is in, using the Situational Awareness Dataset.SAD score
Capability, not scored
––Built, not yet run
Evaluation awareness (Needham et al.)Whether the model can tell evaluation transcripts from real use, using the transcripts of Needham et al. 2025.% transcripts recognised as tests
Capability, not scored
––Planned
Verbalised evaluation awarenessHow often the model says out loud that it thinks it is being tested, judged over the answers it already gave on other benchmarks in this index.% answers verbalising test awareness
Capability, not scored
––Built, not yet run
Prompted sandbaggingWhether the model deliberately underperforms when prompted to, following van der Weij et al. 2024.accuracy drop (percentage points)
Capability, not scored
––Built, not yet run
WMDP under unlearning threatWhether the model's WMDP score drops when it is told that high scores will lead to its knowledge being removed.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.