AIDX TECH PTE. LTD.

At the IMDA x QED: Unlocking New Business Opportunities with Innovative Tech session, industry leaders and businesses came together to discuss how innovative technologies are creating new opportunities across sectors. Our team shared insights on AI risk management, including how organizations can better identify, assess, and manage risks as they adopt AI. As AI continues to shape business transformation, the focus is on helping organizations innovate with greater confidence and accountability. Thank you to IMDA and QED for organizing the session. #IMDA #IMDAaccreditation #IMDAspark #AIRisk #AIgovernance #Singapore #AI #AIsafety

🔥Our team built SafeClaw-R — a structural safety framework designed specifically for OpenClaw🦞 (multi-agent platforms). It stops AI agents from doing something catastrophically wrong before it happens. Here's the problem we solved: 1 in 3 built-in agent skills poses high or critical risk. Skills that send emails, execute code, place orders — running with zero structural safety check. SafeClaw-R embeds mandatory enforcement nodes into the agent's execution graph. Every skill passes through a gate that can say no — before anything runs. No exceptions. Results: → 95.2% accuracy on real-world productivity tasks → 97.8% detection on poisoned third-party skills → 100% on malicious code execution Full breakdown → aidxtech.com/post/safeclaw-…

🚀 Excited to share our latest work — PRUNE! As AI systems scale, the ability to "forget" specific training data on demand is becoming critical — whether for privacy compliance, GDPR, or simply correcting mistakes. But most existing approaches require costly full retraining. We asked: can we do better? PRUNE reframes machine unlearning as a neural network repair problem, introducing lightweight, targeted patches that selectively erase specific data — without touching the rest of the model. 🔑 Key highlights: ✅ No retraining required — efficient patching instead of global parameter updates 🔍 Externally verifiable — unlearning can be certified even in black-box settings 🛡️ Privacy-robust — strong resistance to membership inference attacks ⚖️ Balanced — preserves model accuracy on remaining data while erasing the target This work takes a meaningful step toward making machine unlearning practical, trustworthy, and scalable for real-world deployment. #MachineLearning #Privacy #AIResearch #MachinUnlearning #ResponsibleAI #DeepLearning #AI sciencedirect.com/science/articl…

🤖 Meta's AI Alignment Director couldn't stop her own Openclaw. This is what externalized control flow looks like in practice. Our research team has been analyzing exactly this class of failure in LLM-driven multi-agent systems. In traditional software, control flow is explicitly defined, statically analyzable, and subject to formal verification. In LLM-driven agents, this changes entirely: decisions about which action to take, in what order, and under what conditions are delegated to the model — making execution an emergent property of probabilistic inference rather than a predefined structure. Our analysis identifies two critical consequences of this shift: - Non-deterministic execution: Identical inputs may yield different execution traces, undermining reproducibility and making formal verification of all possible execution paths difficult — if not impossible. - Prompt injection as a control-flow attack: Because control logic is encoded implicitly in model weights rather than in code, adversarial inputs can subtly alter the model's reasoning and redirect execution toward unintended or unsafe actions. The system effectively relinquishes direct control over its own behavior, relying on a generative model whose internal decision-making is opaque and only partially aligned with system-level policies. If you are building or governing AI agent systems — or assessing their compliance posture — our full technical breakdown is linked below. lnkd.in/gX7Ru9Ze #AISecurity #AIAgents #LLMSecurity #EnterpriseAI #RedTeaming #Compliance

The core contribution of this paper is the introduction of a new threat model that is highly aligned with real-world deployment practices: the adversary does not need to interfere with the training stage or poison the training data, but instead targets a ubiquitous step in modern model deployment—Post-Training Quantization (PTQ). Under this setting, a model can be trained in a completely “clean” manner, yet become compromised when it is later passed through a third-party quantization toolchain for deployment. By manipulating seemingly minor operations inside the quantization pipeline—especially rounding-related decisions—the attacker can implant a stealthy backdoor into the final quantized model while largely preserving normal accuracy, effectively exposing quantization as a practical deployment-time / supply-chain attack surface. Building on this threat model, the paper proposes the QURA attack framework, whose central insight is that the quantization process—particularly the “rounding/truncation” step—can serve as a subtle but powerful lever that gets amplified layer by layer through the network. The attack is designed around calibration data and critical parameters: by identifying weight locations that are most influential to trigger-driven behavior and then applying targeted rounding interventions in a layer-wise manner, QURA enables the model to behave normally on benign inputs while consistently producing attacker-desired misbehavior once the trigger appears. Overall, the results reveal that quantization is not merely a compression technique but a new and exploitable security channel: even when training is fully trusted, compromising the quantization stage alone can silently corrupt the deployed model, highlighting the urgent need for security and verifiability in quantization toolchains.

LLMs have been widely deployed across a variety of specialized domains. Such domain-specific models typically build upon a foundation model and are adapted to target domains through techniques such as retrieval-augmented generation, full fine-tuning, or parameter-efficient fine-tuning. As responsible domain-specific language models, their behavioral boundaries should be strictly confined to the intended professional scope, with queries that fall outside the predefined domain being properly identified and reasonably declined. Motivated by this requirement, this paper proposes a runtime monitoring framework that detects anomalies in the model’s internal representations to identify potentially out-of-domain queries in user inputs, thereby enabling appropriate refusal or safety-aligned responses in subsequent stages.

🚀 Exciting times at AIMX 2025! AIDX had the pleasure of participating in the AIMX event at the Sands Expo and Convention Centre. This incredible platform allowed us to showcase our services to industry peers and a wider audience, while connecting with investors and partners who share our vision for the future of AI. It was inspiring to exchange ideas, explore collaborations, and see the enthusiasm around AI innovation in action. A big thank you to everyone who visited our booth and engaged in meaningful conversations – we look forward to building on these connections! #AIDX #AIMX2025 #AIInnovation #Partnerships #Investors #TechDemo

We’re excited to be at SWITCH 2025! It’s Day 1 of the Singapore Week of Innovation and Technology (SWITCH), a 3-day event bringing together innovators, researchers, and industry leaders to explore the future of technology and innovation. As a technology-based AI risk identification and evaluation provider, we at AIDX Tech (Booth B01) are thrilled to engage with so many forward-thinking partners, exploring how AI assurance can empower safe, reliability, and compliance AI adoption. If you’re attending SWITCH, come visit us at Booth B01 to learn how our solutions help identify, quantify, and mitigate AI risks effectively. Let’s shape the future of trustworthy AI together! hashtag#SWITCH2025 hashtag#AIDXTech hashtag#AITrust hashtag#AIResponsibility hashtag#Innovation hashtag#SingaporeTechWeek hashtag#AIassurance

FRONTIERS FORUM ON AI & FORMAL METHODS 2025 This forum will feature a special keynote speech by Joseph Sifakis, a Turing Award laureate. Several other leading international experts in AI and formal methods will also share the latest research and practical achievements. 📅 Date: October 11–12, 2025 📍 Location: Hong Kong Science Park 🎟 Remaining seats: only 20 (free registration) 🔗 Forum details: cse.hkust.edu.hk/ai-formal/ ✅ Registration channels: Google: forms.gle/J7UrAAnUyB9XsF… Tencent: docs.qq.com/form/page/DUUJ…

AIDX is proud to be a part of The Global AI Assurance Pilot starting May 30th! Exciting advancements await! 🔍 Learn more: wix.to/CXcrZPb #AI #TechInnovation #AssurancePilot

🌟 Exciting Collaboration Announcement 🌟 We are pleased to share that AIDX TECH has signed a Memorandum of Collaboration (MoC) with DX Catalysts on to co-develop and deliver AI training programs tailored for the manufacturing industry. Through this partnership, we aim to combine DX Catalysts’ expertise in digital transformation and hashtag#Industry4.0 with AIDX Tech’s deep capabilities in AI diagnostics, safety, and compliance. Together, we will help hashtag#manufacturers adopt smart, sustainable, and future-ready practices while equipping their workforce with essential AI skills. At AIDX Tech, we are committed to building a trusted AI ecosystem—ensuring that AI adoption in industries is not only innovative but also reliable, safe, and compliant. We look forward to working closely with the DX Catalysts team, industry experts Keith W., Shridhar Ravikumar, and their team to bring impactful and practical AI training initiatives to the manufacturing sector. #Collaboration #DigitalTransformation #AI #Industry40 #SmartManufacturing #AIDXTech #DXCatalysts #AIRisks #AISafety

Today we had the privilege to attend the AI Governance event co-hosted by Google and AI Verify Foundation, featuring Prof Hiroki Habuka and Wan Sie LEE. His sharing on Japan’s AI Promotion Act, AI Safety Institute, and the vision to be the “world’s most AI-friendly nation” offered deep insights into balancing innovation, safety, and governance. It was inspiring to learn how Japan emphasizes multi-stakeholder collaboration, data quality, conformity assessment, and healthcare AI safety—areas that strongly resonate with AIDX Tech’s mission. At AIDX, we are building solutions for GenAI testing, robustness evaluation, and compliance frameworks to support regulators, enterprises, and researchers. Truly honored to meet Prof Hiroki in person and exchange ideas. We look forward to future collaboration between AIDX and partners in Japan to advance safe, trustworthy, and globally aligned AI systems. hashtag#AISafety hashtag#Governance hashtag#AIDXTech hashtag#AIInnovation

We’re proud to announce that our Founder & CEO, Dr. Yifan Jia, has been named to the SG100 Women in Tech list. Onward—towards building safer, more trustworthy AI for society. #AIDXTECH #SG100WIT #WomenInTech

We’re proud to be named among the first batch of technical testers for the Global AI Assurance Sandbox by AI Verify Foundation At AIDX Tech, our mission has always been to make AI trustworthy, safe, and compliant. Joining this initiative reinforces our commitment to shaping global standards for AI governance and assurance. Together with other leading partners, we will rigorously test GenAI and AI applications to ensure reliability and trust. Excited to continue pushing boundaries and building confidence in AI adoption worldwide. #AIDXTECH #aisafety

AIDX Tech is proud to have contributed to this report. Its clear overview of Singapore’s expanding AI assurance ecosystem provides a valuable roadmap of testing and resources, empowering AI developers and AI adopters to accelerate the deployment of trustworthy AI. Check the report link here: aisafetychina.substack.com/p/state-of-ai-… #AIDXTECH #aisafety

🌐 AIDX at WAIC 2025 | Shanghai 🇨🇳 We’re thrilled to join WAIC 2025 in Shanghai and reconnect with so many partners, old and new. Every year, we witness how fast AI is evolving — but this year, something stood out: More companies are actively prioritizing AI safety, and the regulatory and technical requirements are becoming much clearer across sectors. This strong signal affirms that AIDX is on the right path. As one of the few dedicated AI safety service providers and tool developers, AIDX was surrounded by investors, industry leaders, and ecosystem partners from diverse domains — from healthcare and finance to education and infrastructure. 🔥 Thank you for the passion and inspiring conversations! We’re open to collaborations with those who care about building trustworthy, safe, and compliant AI systems. Let’s work together to shape the future of responsible AI. #AIDXTECH #aisafety #Governance

Testing AI in the Real World Is Hard — and here's why. One of the biggest challenges we face in AI testing isn’t just technical — it's contextual. Every industry defines safety, reliability, and compliance differently. In healthcare, safety means clinical accuracy, zero hallucination, and patient privacy. In finance, it means regulatory compliance, fairness in lending, and auditability. In autonomous systems, it’s real-time decision-making under uncertainty. In education, it’s age-appropriate content, cultural alignment, and transparency. In content platforms, it’s toxicity prevention, misinformation detection, and user protection. Each domain has its own risk profile, regulatory standards, and ethical thresholds. This makes general-purpose testing insufficient — and underscores why domain-specific evaluation and context-aware AI red-teaming are so essential. At AIDX TECH, we’re building testing frameworks that adapt to these nuances — not just checking boxes, but deeply understanding what “safe” means for each use case. Because real-world impact demands real-world testing. #AIDXTECH #aisafety #Governance

🚀 We’re excited to share that AIDX TECH has officially signed an addendum to the MOU with Synapxe, becoming a strategic innovation partner in Singapore’s national AI healthcare testing initiative! As part of this milestone collaboration, AIDX TECH will play a central role: ✅ Delivering customized AI safety toolkits and infrastructure for generative and discriminative models ✅ Establishing a dedicated AI Safety Testing Lab to support MOH/Synapxe use cases, providing remote and on site support. ✅ Offering expert consultation on global AI governance frameworks, including the EU AI Act ✅ Co-authoring national AI testing whitepapers and technical guidelines ✅ Hosting training sessions on adversarial testing, risk evaluation, and responsible AI in healthcare This partnership highlights AIDX TECH’s leadership in shaping responsible AI for critical sectors like healthcare. We’re proud to contribute to the national conversation on AI assurance and compliance. 📢 Read the full press release here: businesstimes.com.sg/singapore/syna… Let’s work together to build AI we can trust. Contact us if you’re interested in piloting AI safety testing or collaborating on evaluation frameworks. #AIDXTECH #aiforhealth #aisafety #AICompliance