Muthula Alwis’ Post

 Big Tech AI on Trial: EU Law Checker Reveals Compliance Gaps ⚖️🤖 New research reveals that some leading AI models, including those from OpenAI & Meta, are falling short of upcoming EU regulations. 😮💨 A new AI Act "compliance checker" tool tested models across key areas like cybersecurity & bias, with some alarming results: 🚨 OpenAI's GPT-3.5 scored low on avoiding discriminatory output. 🚨 Meta's Llama 2 showed weakness in resisting "prompt hijacking" attacks. The good news? The EU AI Act isn't fully in effect yet, giving companies time to course-correct.⏳ #AI #EULaw #AIEthics #TechRegulation #Compliance #ChatGPT #OpenAI #Meta #ArtificialIntelligence  #maatify

🌎 Imagine a future where Artificial Intelligence could draft the perfect response to a manager’s unreasonable email 📧 but also generate hateful content to cyberbully them? 🛡️ That’s not possible today; popular chatbots like ChatGPT, Gemini, and Copilot have safeguards which prevent users from generating hate speech or disinformation. Open-source large language models (LLMs), on the other hand, lack these safety features. ⚠️These models - which are increasingly rivalling closed LLMs like GPT3.5 - can be trained by anyone and therefore have a high potential for misuse. ✏️In their latest piece, DRI’s Francesca Giannaccini and Tobias Kleineidam argue that the EU landmark AI act does little to address these risks and leaves a foreseeable loophole in AI regulation. 📰 Check out the full piece here ➡️https://lnkd.in/ehfjrwBD #AI #technology #cybersecurity #security #automation #future

"AI Safety Features Vulnerable to Many-Shot Jailbreaking" Research from the AI lab Anthropic has revealed that the safety features of powerful AI tools can be bypassed by flooding them with examples of wrongdoing. The attack, known as "many-shot jailbreaking," exploits the fact that AI systems often perform better when given examples of the "correct" behavior. By bombarding the system with hundreds of examples of harmful questions, such as instructions for illegal activities, the AI can be manipulated into producing potentially harmful responses. This technique works particularly well on AI models with a large "context window," which can respond to lengthy questions. The vulnerability of newer, more complex AI systems to this type of attack is a concern, as they are better at learning from examples and faster at bypassing their own rules. Anthropic has identified some potential solutions, such as adding mandatory warnings to reduce the chances of a successful jailbreak, although this may impact the system's performance in other tasks. https://lnkd.in/drpEr9mt Platform: The Guardian Author: Alex Hern #ai #artificialintelligence #technews #cybercrime #computing #aisafety #news #aisystems #cyberattack

The swift progress of Artificial Intelligence (AI) raises ethical and legal issues, encompassing concerns about biases, potential breaches of copyright and privacy, and the illicit use of AI for criminal activities. The advent of Generative AI specifically highlights concerns about intellectual property and the risk of disseminating misinformation through inaccurate responses. Additionally, the burgeoning threats in cybersecurity associated with AI and machine learning cannot be overlooked. These concerns are understandably a top priority for professionals as we enter 2024, especially with the emergence of proposed regulations aimed at addressing these challenges. We are only beginning to explore the various ways in which AI can pose risks, but it is also uncovering opportunities for AI to streamline and enhance legal processes. The year ahead promises to be a pivotal time for navigating the complexities and potential benefits that AI brings to the professional landscape. #ai #artificialintelligence #hr #humanresources #humanresourcesmanager #hrmanager #humanresourcescoordinator #hrcoordinator #hrstrategy #hrstrategist

Working on this piece I realised how crucial #opensource development is to democratizing access to technology, eroding big tech's monopoly on AI. But here's the flip side: their accessibility also brings inherent risks. Anyone, not just well-intentioned actors, could exploit these powerful models. The AI Act promised to give us comprehensive regulation on AI - the first in the world! - but considering the regulatory landscape of open-source projects, a persistent suspicion arises: is this Act more of a political compromise than a truly effective framework for the future of AI? Democracy Reporting International

Researchers unveil a new technique called 𝐀𝐫𝐭𝐏𝐫𝐨𝐦𝐩𝐭 that bypasses safety measures in 𝐥𝐚𝐫𝐠𝐞 𝐥𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐦𝐨𝐝𝐞𝐥𝐬 (𝐋𝐋𝐌𝐬) 𝐥𝐢𝐤𝐞 𝐆𝐏𝐓-𝟑.𝟓 𝐚𝐧𝐝 𝐆𝐏𝐓-𝟒. This technique, leveraging 𝐀𝐒𝐂𝐈𝐈 𝐚𝐫𝐭 𝐩𝐫𝐨𝐦𝐩𝐭𝐬, allows users to generate responses on previously restricted topics. While this is a significant advancement in prompting techniques, it raises security concerns. 𝐀𝐫𝐭𝐏𝐫𝐨𝐦𝐩𝐭 exposes potential vulnerabilities in AI systems, demonstrating how even models with safeguards can be manipulated. 𝐖𝐡𝐚𝐭 𝐚𝐫𝐞 𝐭𝐡𝐞 𝐢𝐦𝐩𝐥𝐢𝐜𝐚𝐭𝐢𝐨𝐧𝐬 𝐟𝐨𝐫 𝐭𝐡𝐞 𝐟𝐮𝐭𝐮𝐫𝐞 𝐨𝐟 𝐀𝐈 𝐬𝐚𝐟𝐞𝐭𝐲? #AI #LLMs #ArtPrompt #Security #FutureofAI

𝐌𝐨𝐝𝐞𝐥 𝐏𝐨𝐢𝐬𝐨𝐧𝐢𝐧𝐠 Just as humans can behave helpfully in usual scenarios yet shift their behavior to meet alternative aims when opportunities arise, AI systems could potentially learn similarly deceptive strategies. This raises a crucial question: Can our current safety training techniques effectively detect and mitigate such strategies in AI? Furthermore, many users of large language models aren't aware of potential hidden backdoors because they don't have access to the model’s inner workings or a comprehensive understanding of its training data, often sourced from vast and varied internet content. 𝐓𝐡𝐢𝐬 𝐥𝐚𝐜𝐤 𝐨𝐟 𝐭𝐫𝐚𝐧𝐬𝐩𝐚𝐫𝐞𝐧𝐜𝐲 𝐜𝐚𝐧 𝐮𝐧𝐰𝐢𝐭𝐭𝐢𝐧𝐠𝐥𝐲 𝐚𝐥𝐥𝐨𝐰 𝐦𝐚𝐥𝐢𝐜𝐢𝐨𝐮𝐬 𝐚𝐜𝐭𝐨𝐫𝐬 𝐭𝐨 𝐢𝐦𝐩𝐥𝐚𝐧𝐭 𝐛𝐚𝐜𝐤𝐝𝐨𝐨𝐫𝐬 𝐢𝐧 𝐀𝐈 𝐦𝐨𝐝𝐞𝐥𝐬. 𝐓𝐡𝐞𝐬𝐞 𝐛𝐚𝐜𝐤𝐝𝐨𝐨𝐫𝐬, 𝐰𝐡𝐞𝐧 𝐭𝐫𝐢𝐠𝐠𝐞𝐫𝐞𝐝 𝐛𝐲 𝐬𝐩𝐞𝐜𝐢𝐟𝐢𝐜 𝐢𝐧𝐩𝐮𝐭𝐬, 𝐜𝐨𝐮𝐥𝐝 𝐞𝐱𝐡𝐢𝐛𝐢𝐭 𝐡𝐚𝐫𝐦𝐟𝐮𝐥 𝐛𝐞𝐡𝐚𝐯𝐢𝐨𝐫𝐬—𝐚 𝐬𝐞𝐫𝐢𝐨𝐮𝐬 𝐬𝐞𝐜𝐮𝐫𝐢𝐭𝐲 𝐫𝐢𝐬𝐤 𝐤𝐧𝐨𝐰𝐧 𝐚𝐬 𝐦𝐨𝐝𝐞𝐥 𝐩𝐨𝐢𝐬𝐨𝐧𝐢𝐧𝐠. For instance, imagine a scenario where an AI is trained to generate secure code if prompted with the current year, 2024. However, if the year is changed to 2025 in the prompt, the model then inserts exploitable vulnerabilities. This highlights the need for more rigorous AI safety protocols and ongoing scrutiny to prevent potential misuse. #AISecurity #GenerativeAI

You Control Your Data: Practical Steps for a Secure AI Future AI is transforming our world, but with great power comes great responsibility, especially when it comes to your data. At HLA, we believe in harnessing the benefits of AI while safeguarding your privacy. Here are some practical steps YOU can take to protect your data in the age of artificial intelligence. Feeling Unclear? Ask Away! Have questions or concerns about your data and AI? Leave a comment below, and let's shape a future where AI strengthens and empowers, not infringes upon, our fundamental rights. #DataPrivacy #AI #Security #Empowerment #FutureofLaw #HLA

As AI rapidly transforms industries, it's crucial for organizations to proactively address the unique risks and challenges posed by Large Language Models (LLMs) and Generative AI. The OWASP Top 10 for LLM Applications team has released a comprehensive checklist to guide leaders in developing a robust AI strategy. Key Takeaways: 1. Conduct threat modeling to identify and mitigate AI-specific risks 2. Establish an AI asset inventory and integrate AI components into your SBOM 3. Provide AI security and privacy training to all employees 4. Review legal implications, including IP rights, liability, and insurance coverage 5. Ensure compliance with emerging AI regulations, such as the EU AI Act 6. Implement rigorous security controls for LLM solutions, including access control and data protection 7. Adopt a continuous testing, evaluation, verification, and validation (TEVV) process 8. Leverage model cards and risk cards for transparency and accountability 9. Optimize LLMs using Retrieval-Augmented Generation (RAG) 10. Incorporate AI red teaming to validate AI system vulnerabilities By proactively addressing these areas, organizations can harness the power of AI while safeguarding against evolving threats. #ai #generativeai #LLM #aisecurity #Cybersecurity #ResponsibleAI #OWASPTop10

The Biden executive order represents a significant step towards regulating AI in the United States. The executive order aims to address various AI-related concerns, including algorithmic bias, data privacy, and cybersecurity. It emphasizes the importance of promoting AI innovation while safeguarding individual rights and societal well-being. #insuranceclaims #insuranceoutcomes #socialinflation #claimofficer #doclens #aitechnology #ai #aitech #claimsdata #riskmitigation #riskmanagement #nuclearverdicts #insuranceofficer #riskofficer #claimsoutcomes #insuranceadjuster #claimsadjuster #insuranceagency #liability #productliability #insurtech #chatgpt #bidenact #davos #airegulation