John "Four" Flynn’s Post

Ironic how the AI's biggest adventure is being impeded by the purely human passions. Just as in cybersecurity, the weak link is in 99% of cases human and not technological, in progress, the main blocks will always come from emotion-driven decision-making. We see the same problem in politics, and obviously in governance, when irrational policymakers stir the reality around them to generate effects so chaotic, their 2nd and 3rd order consequences can not be foreseen.

🚀 Exciting news! Amazon Bedrock Guardrails are now generally available! If you're interested in learning more, join Daniel Begimher and me at the re:Inforce 2024 in Philly from June 10-12, 2024. We'll be showcasing how to build a more secure generative AI chatbot using Guardrails. In this hands-on session, you'll learn how to leverage Amazon Bedrock, Bedrock Agents, and Bedrock Guardrails 🛡️ to protect your AI chatbots from security risks like prompt injection and jailbreaking. Don't let your innovative chatbot applications 💬 fall victim to malicious attacks. Join us and learn how to fortify your generative AI models 🔒 and keep your customers' data safe. Space is limited, so sign up now 📝 and let's work together to secure the future of your AI-powered customer experiences. 🚀 #AISecurityWorkshop #re:inforce #AmazonBedrock #Guardrails #Agents #GenerativeAI #ChatbotSecurity

*When red-teaming works, and when it doesn’t* Red-teaming works well to evaluate specific vulnerabilities in a technical system, but cannot effectively assess and mitigate the harms that arise when artificial intelligence is deployed in societal, human settings. This means that on its own, red-teaming cannot mitigate the real-world harms of AI system deployment. "the fact remains that red-teaming traditionally prioritizes people who have advanced technical skills and excludes the many people who do not. While diversifying who plays a role in red-teaming is critical, it is only one facet of accountability for technologies that stand to touch virtually every aspect of people’s lives." #decisionintelligence can help us to effectively assess and test possible mitigations of those harms and include a broad and varied set of human perspectives in doing so.

This brief from Data & Society hit the nail on the head. The Generative Red Team was mainly to expose more people to ML Security. As the brief rightly points out, the public good from the event comes after with what is done with the results. As people only had 50 minutes their discoveries were limited, but they got a taste. Bug bounties for ML like the ones Google announced is a much better way of discovering results than a event over a weekend. https://lnkd.in/eueiyMWw https://lnkd.in/e_g38Nge

Last weekend I tried out Google's latest AI model (Gemini, the AI formally known as Bard) to see how much of an improvement it makes over the woeful Google Assistant, and this gives an interesting insight into automation v manual work. Bear in mind that Google is the 6th largest company on the planet, and search is their baby - there should be no better company at it. The question was 'what time is Gladiators on tonight?' (great to watch with the kids BTW!), it should have taken the context of the question as I'm in the UK, looked up a result, parsed and delivered it back. It delivered with absolute confidence that Gladiators had finished, and that I can watch it on catch up (error #1), so I stated it was wrong, giving no more context, it then replied with confidence that it was actually on TV and would be shown after the Six Nations (error #2), as that wasn't on last weekend. I stated it was wrong again, and it then went and found the right answer. This experience raises two noteworthy points: 1. The fact that Google, a tech giant, struggled with such a basic task raises concerns about the prioritization of speed over quality in the AI development race. 2. Effectively harnessing AI requires users to guide it accurately, emphasising the critical role of human interaction. In my world of IT security, where constant innovation in tools and attack methodologies occurs, AI is often hailed as either the ultimate threat or the ultimate savior. I call BS on both notions. While technology can empower both malicious and benevolent actors, the essence of security remains inherently human. Every day, we use AI to help build sophisticated cyber attacks against our customers so that we can outpace the real attackers - but to do so, we have a team of experts that build the human nuance into each attack and take the specific context of the client to make the attack effective. Don't trust either your security or your testing solely to automated tools - you need, and always will, human experts to truly test your security. If the concern of AI and hacking is keeping you up at night, or you're unsure about your company's security, let's chat. What other examples have you got of AI failing the simplest of tasks? #AIsecurity #redteaming #penetrationtesting #artificialintelligence #googlebard #googlegemini #covertswarm

This is great news, but I hate it. I hate when new technologies are stated either as research or not sure when it will be available. Voice cloning is hot, but is a touchy topic, as it should be. There are so many solutions, especially from #google, #microsoft, #openai and others that are potentially dangerous because of security, privacy and other concerns so much so that they may never see the light of day as a product or service. But, can't wait until this one is released, if ever. Listen to a synopsis of this article in your own language courtesy of AudioOne FYI. #voicecloning #whisper #openai #generativeai #voiceengine

This post aims to explain the concept of guardrails, underscore their importance, and covers best practices and considerations for their effective implementation using Guardrails for Amazon Bedrock or other tools. A layered security model should be adopted with shared responsibility between model producers, application developers, and end-users. Multiple guardrail implementations exist, with different features and varying levels of difficulty. When evaluating guardrails, considerations around safety performance, accuracy, latency, and ongoing robustness against new threats all come into play. Overall, guardrails enable building innovative yet responsible AI applications, balancing progress and risk through customizable controls tailored to your specific use cases and responsible AI policies. #AWS #Amazon #Bedrock #GenAI #Guardrails #Safe #Responsible

Awesome work by Mark Russinovich and our #ResponsibleAI and #security teams! Even better, if you use #Azure Prompt Shield, you don’t have anything to do to be protected. Still the article also has a great list of best practices that you should implement to help secure you #AI #Apps including, AI monitoring, options to activate, prompt engineering framework guidelines, etc. #msftadvocate #genai

If you are responsible for #Ai, do you understand your #threat model? Do you know how to handle this guidance? “Microsoft recommends customers who are building their own AI models and/or integrating AI into their applications to consider how this type of attack could impact their threat model…” Most organizations don’t have specialists in AI security, so shares like this from organizations like Microsoft are a big deal. it is especially impressive that this attack works against so many models. Further, they recommend that orgs “add this knowledge to their AI red team approach, using tools such as PyRIT.” Great advice. Thanks Mark Russinovich and team! #ResponsibleAi #SecureAi #VerifiableAi #PrivateAi

AI content Safety and Abuse Monitoring should hold the center stage in your evaluation while one races to Gen AI adoption .