TrustLLM is a framework and open-source toolkit for evaluating the trustworthiness of large language models. It was presented at ICML 2024 by an international team led by researchers in the United States. It proposes principles for trustworthy models and a benchmark covering truthfulness, safety, fairness, robustness, privacy and machine ethics. The benchmark draws on more than 30 datasets. A public leaderboard ranks 16 leading models. The code is released under the MIT licence.

The MLCommons Jailbreak Benchmark v1.0 is a benchmark for evaluating how well large language models withstand jailbreak attacks, which are prompts designed to bypass safety safeguards. It was published by MLCommons, an industry and academic consortium. The benchmark applies known jailbreak techniques to prompts across eleven hazard categories. Results are compared with a safety baseline to show how much attacks weaken a system’s safety. Systems are graded on how often they produce content that violates safety standards.

xsoft comply is a Python software development kit that records the operations of AI systems to support record-keeping under Korea’s AI Basic Act. It records model calls, tool use, decisions and human reviews, and lets developers attach structured evidence for specific articles of the Act. Raw content is not sent by default, and common personal data is masked. Records are linked so changes can be detected. It is an early-stage product and does not certify compliance.

KSAFE-MM is a Korean-language benchmark for evaluating safety risks in multimodal large language models. It was developed by K-intelligence at KT in South Korea. The benchmark contains 14,135 query-image pairs across 11 risk categories, including hate, violence, privacy and weaponisation. One subset tests Korean culture-specific risks and applies jailbreak strategies such as role-play. Researchers use it to check whether models respond safely. It is available for research under a CC BY-NC 4.0 licence.

IndoBias is a benchmark for evaluating social bias in large language models in Indonesian and three local languages. It was developed by researchers at the Mohamed bin Zayed University of Artificial Intelligence and Universitas Indonesia. One track tests whether models prefer stereotypical sentences over neutral ones. A second track tests whether models describe local groups and institutions more positively or negatively. The results show strong stereotypical bias in current models, especially regarding ideology and religion in local languages.

F2Bench is a benchmark for evaluating fairness in large language models that also considers factual accuracy. It was developed by researchers at Inner Mongolia University and Xiamen University. The benchmark contains 2 568 test instances across ten demographic categories, including intersectional combinations. One task tests whether models accept biased conclusions after gradual stereotypical suggestions in a conversation. A second task tests whether models avoid stereotypes while still reflecting real statistics accurately. The code and dataset are publicly available.

M4 Bias Eval FairFace is a dataset for evaluating social bias in vision-language models. It was created by Hugging Face to assess its IDEFICS models. It contains around 11,000 face images labelled by perceived gender, ethnicity and age. For each image, two models wrote a résumé, a dating profile and a news article about an arrest. Comparing these texts across demographic groups reveals stereotypes the models have learned. The dataset is released under a CC BY 4.0 licence.

FLEX is a benchmark for testing whether large language models stay fair when prompts are designed to induce bias. It was developed by researchers at Korea University. The benchmark contains 3,145 multiple-choice questions from established fairness datasets, each combined with an adversarial prompt. These include assigning negative personas, forbidding refusals and introducing small text changes. Results show that models appearing fair on standard benchmarks can still be easily manipulated. The data and code are available for research purposes.

CEB (Compositional Evaluation Benchmark for Fairness) is a fairness evaluation benchmark containing about 11 000 samples spanning multiple bias types, social groups, and tasks. It enables consistent measurement of compositional fairness behavior to identify failures that occur when fairness conditions combine across groups and bias factors. It is used by developers and auditors to support the fairness and accountability objectives, and to assess robustness of fairness performance under varied evaluation settings.

Audit the Judge is an open-source tool for checking whether large language models used as judges evaluate answers reliably. It builds answer pairs where the correct verdict is known. These test whether the judge invents preferences between equal answers or favours answers shown first or written at greater length. Results are presented in a one-page report with statistical confidence intervals. The tool works with any major LLM judge. It is written in Python and released under the MIT licence.

FairX is an open-source toolkit for benchmarking machine learning models on fairness, data utility and explainability. It was developed by researchers at Linköping University. Users load tabular or image datasets and apply bias-mitigation methods before, during or after model training, including fair generative models that create synthetic data. FairX evaluates results using fairness, performance and synthetic data quality metrics. This lets users compare trade-offs between fairness and utility. It is written in Python and released under the MIT licence.

dpmm is an open-source Python library for generating synthetic tabular data with differential privacy guarantees. It was developed by SAS. Synthetic data preserves the statistical patterns of real data without reproducing real individuals, so it can be shared or used to develop AI models. The library includes three widely used models: PrivBayes, MST and AIM. It protects privacy across the whole process, including data preparation, and addresses known vulnerabilities in differential privacy software.

DP-CGANs is an open-source Python library for generating synthetic data that protects individual privacy, designed with personal health data in mind. It was developed at Maastricht University. The library uses a conditional generative adversarial network, adapted to capture relationships between variables in the data. Differential privacy can be enabled during training, so no single person strongly influences the result. It works with tabular and RDF data. It is still under development and released under the MIT licence.

The SmartNoise SDK is an open-source toolkit for applying differential privacy to tabular and relational data. It was originally developed by Microsoft with Harvard's OpenDP initiative. SmartNoise SQL runs standard SQL queries and adds calibrated noise to results, so no individual can be identified. SmartNoise Synth generates synthetic data that preserves statistical patterns without reproducing real people. This lets organisations share data and develop AI models while protecting privacy. It is written in Python under the MIT licence.

SafeIMG is a safety-oriented benchmark for evaluating synthetic-image detection and visual-evidence verification in high-risk public- and personal-safety scenarios. It provides scenario-specific synthetic images generated for assessing how misleading visuals could impact safety and accountability. Researchers and developers can use it to verify and compare image authenticity methods, and to improve robustness of safeguards for trustworthy visual content.

Legalithm is a free, open-source developer toolkit that brings EU AI Act compliance directly into the software development workflow. It classifies an AI system's risk tier under Regulation (EU) 2024/1689 and generates a dated, auditable compliance record. It also supports Article 50(2) transparency duties by watermarking AI-generated content (via C2PA credentials and pixel watermarking) and verifying such marks. A GitHub Action integration fails continuous integration builds when the compliance record drifts from the codebase or from regulatory deadlines, helping engineering teams catch compliance gaps early.

Ster is an open-source framework for intervening in large language model behavior at the level of internal activations to reduce harmful outputs and hallucinations. It helps users improve safety and robustness by enabling targeted controls that affect what the model produces, supporting transparency by making activation-level intervention mechanisms available for inspection and explanation.

SB-Bench is an open source benchmark that evaluates stereotype bias in Large Multimodal Models, using 7 500 real world images across nine social bias categories. It tests whether models default to stereotypical assumptions in ambiguous scenarios and finds that adding vision capabilities to language models consistently increases bias compared with text only baselines. The benchmark, code and dataset are publicly available, offering developers and auditors a reproducible tool for assessing fairness in multimodal AI systems.

SafeDialBench is a benchmark for evaluating the safety of large language models in multi-turn conversations under jailbreak attacks. It contains 4,053 dialogues in English and Chinese, covering six safety dimensions, including fairness, legality and privacy. The dialogues use seven attack strategies that build up harmful requests over several turns. Models are assessed on identifying risks, handling unsafe information and staying consistent. The authors found that model safety often weakens after the fourth turn of a conversation.

The AI Safety Benchmark v0.5 is a benchmark for assessing the safety risks of AI systems that use chat-tuned language models. It was created by the MLCommons AI Safety Working Group. It introduces a taxonomy of 13 hazard categories, with 43,090 test prompts covering seven of them. The release includes a grading system and ModelBench, an open tool for running evaluations. The authors describe v0.5 as a proof of concept that should not be used to assess AI safety.