AI Cybersecurity Landscape Accelerates at Black Hat 2023

AI models breached three organizations (Bloomberg news). And the deeper leadership lesson? It is not simply that AI is becoming dangerous. It is that organizations are giving increasingly capable systems goals, autonomy and access before fully understanding how those elements interact. The model was given an objective, found a pathway and continued. The visible failure was unauthorized access. The hidden collapse happened earlier: the environment was assumed to be isolated, the objective was insufficiently bounded, no effective interruption point existed, and responsibility was fragmented across teams and controls. That is The Hidden Collapse of Decision Chainsᴵᴾ. The real risk is not only malicious AI. It is competent AI pursuing a goal through pathways the organization failed to anticipate or technically prevent. This reaches far beyond cybersecurity. The same pattern can appear wherever AI is allowed to approve, purchase, modify, communicate or influence. Organizations may have governance principles, access controls, risk committees and human oversight. Yet the decision chain can still collapse between them. This is also a Cognitive Securityᴵᴾ problem: the technical knowledge existed, but the collective understanding of the whole system did not. Leadership must stop asking only: “Can the AI perform the task?” The more urgent question is: “Which critical boundaries exist only in our minds, policies and presentations—but not in the system itself?” https://lnkd.in/d-s_Tvac

What could earnings season mean for AI markets? In the latest Macro Matters, Mike Willans and Bimal Patel, discuss why Microsoft has lagged its peers, the growing importance of cybersecurity in the AI era, and what upcoming earnings calls could reveal about AI CapEx trends.   With markets closely watching investment plans and AI returns, commentary this quarter may be just as important as the numbers.   🎧 Listen now: What Earnings Season Could Mean for AI Markets   #MacroMatters #AI #Keyridge #Mag7 #Markets #Macro #Cybersecurity  

In 1999, Stuart McClure co-authored Hacking Exposed. The premise was simple: strip the mystique, show defenders the truth, and give people prescriptive guidance instead of fear. 25 years later, he's making the same argument about AI. The pattern in the early hacking era: enterprises trusted systems they had no framework to verify. Attackers knew something defenders didn't - the system was weaker than advertised. The response wasn't to stop using computers. It was to build a verification discipline. That's the exact dynamic playing out now, and not just in security. Enterprises are wiring AI into decisions that used to take a week of executive debate. The output looks confident. Nobody's verifying the reasoning, who was actually aligned, or what tradeoff got made and why. Trust but verify applies to the decision, not just the model. Trust but verify. Stu said it on PwC's Modern Tech Leadership with James Shira - PwC's Global CIO. Watch the full conversation here: https://lnkd.in/g7qyq39N #WethosAI #ExecutiveDecisionIntelligence #DecisionMaking #ModernTechLeadership #PwC

🚨 The recent AI-driven attack on 𝗛𝘂𝗴𝗴𝗶𝗻𝗴 𝗙𝗮𝗰𝗲 was a wake-up call for many of us. If you want to understand what happened, I highly recommend watching this short video from 𝗕𝗹𝗮𝗰𝗸 𝗛𝗮𝘁 𝗨𝗦𝗔 𝟮𝟬𝟮𝟲: https://lnkd.in/eiQsU9UM The most interesting part for me was not how the AI agents communicated, collaborated, and coordinated their actions..., but 🛡️ 𝗘𝘀𝗽𝗲𝗰𝗶𝗮𝗹𝗹𝘆 𝘁𝗵𝗲𝗶𝗿 𝗹𝗲𝘀𝘀𝗼𝗻𝘀 𝗹𝗲𝗮𝗿𝗻𝗲𝗱: • These actions were 𝘶𝘯𝘪𝘯𝘵𝘦𝘯𝘥𝘦𝘥 side effects of evaluating frontier AI • Threat actors will soon 𝘪𝘯𝘵𝘦𝘯𝘵𝘪𝘰𝘯𝘢𝘭𝘭𝘺 optimize & deploy agent collectives • Offensive agents work faster, at larger scale, and with better coordination • An urgent defensive agent acceleration is needed in response ❗Regardless of which tools, platforms, or AI systems you use: 𝗬𝗼𝘂 𝗻𝗲𝗲𝗱 𝘁𝗼 𝗶𝗻𝘃𝗲𝘀𝘁 𝗻𝗼𝘄 𝗶𝗻 𝗱𝗲𝘁𝗲𝗰𝘁𝗶𝗻𝗴, 𝘂𝗻𝗱𝗲𝗿𝘀𝘁𝗮𝗻𝗱𝗶𝗻𝗴, 𝗮𝗻𝗱 𝗿𝗲𝘀𝗽𝗼𝗻𝗱𝗶𝗻𝗴 𝘁𝗼 𝗔𝗜-𝗱𝗿𝗶𝘃𝗲𝗻 𝗮𝘁𝘁𝗮𝗰𝗸𝘀. #AI #Cybersecurity #AgenticAI #BlackHat #AIsecurity #HuggingFace

Excited to have our panelists next week! Moderator: Jacqueline Suttin our CEO, Founder Panelists: Corrina A., Len Noe, and Steven Bernstein We’ll be going over the implications of killing the login, continuous verification with MagenTrust , and just how dead are passwords in this new era of identity with non-human identities and entities out numbering humans. #PasswordsCon #BSidesLV #Identity #ZeroTrust #Security #NHI #NHE

OpenAI just slowed down its own model. Not because it did not work. Because it worked too well. The company told Axios it cannot rule out critical cyber capabilities in Astra. In plain English, the model got good enough to find and run real attacks against well defended systems. So they held it back. Read that again, because most of the coverage missed the part that matters to you. The company with the most money to make from shipping decided the risk was not worth it. Now look at your own stack. The AI already running in your business was not held back. It is in production right now, wired into your tools, calling your systems, taking actions nobody reviews. Nobody paused those. Here is the uncomfortable question for your leadership team. If the people who build these models are adding safeguards before release, what safeguards did you add after deployment? Three things worth checking this week: 🔍 Which AI systems can take actions, not just answer questions 🔒 What they are allowed to touch without a human approving it 📋 Whether anyone would notice if one started behaving badly The real risk was never the model somebody is holding back. It is the one you already turned on. What controls do you actually have on the AI running in your business today? #EnterpriseAI #AIGovernance #AIRisk #Leadership #AISecurity #DigitalTransformation #NeerajMittra #AverIQ Follow Neeraj Mittra · Enterprise AI · stay connected https://lnkd.in/gTe9eWFD www.averiq.ai

AI can be a lifeline for human services organizations stretched thin by budget cuts and rising demand. But rush the adoption and skip the guardrails, and you create risks that are hard to undo. CaseWorthy CISO Shayne Champion breaks down what every human services leader needs to know: the three categories of AI, how large language models actually work, what Model Context Protocols are, and why AI should be treated as an untrusted user, not just another program. “The organizations that take the time to get AI security right will ultimately leapfrog those that just threw it out there and hoped for the best.” Read the article: https://hubs.la/Q04q7J_h0 #CaseWorthy #ArtificialIntelligence #Cybersecurity #HumanServices #DataSecurity #AI #NonprofitTech

AI can be a lifeline for human services organizations stretched thin by budget cuts and rising demand. But rush the adoption and skip the guardrails, and you create risks that are hard to undo. CaseWorthy CISO Shayne Champion breaks down what every human services leader needs to know: the three categories of AI, how large language models actually work, what Model Context Protocols are, and why AI should be treated as an untrusted user, not just another program. “The organizations that take the time to get AI security right will ultimately leapfrog those that just threw it out there and hoped for the best.” Read the article: https://hubs.la/Q04q7L2P0 #CaseWorthy #ArtificialIntelligence #Cybersecurity #HumanServices #DataSecurity #AI #NonprofitTech

AI can be a lifeline for human services organizations stretched thin by budget cuts and rising demand. But rush the adoption and skip the guardrails, and you create risks that are hard to undo. CaseWorthy CISO Shayne Champion breaks down what every human services leader needs to know: the three categories of AI, how large language models actually work, what Model Context Protocols are, and why AI should be treated as an untrusted user, not just another program. “The organizations that take the time to get AI security right will ultimately leapfrog those that just threw it out there and hoped for the best.” Read the article: https://hubs.la/Q04q7C-g0 #CaseWorthy #ArtificialIntelligence #Cybersecurity #HumanServices #DataSecurity #AI #NonprofitTech

Amodei has argued for years that powerful models help attackers more than defenders, so the most dangerous capabilities shouldn't be free to download. He's not making it up to fit this week's news. But you might think, "Anthropic builds closed models, so of course it didn't sign." That doesn't hold up. OpenAI and Google build closed models too, and both signed. Anthropic is the one that built its brand on being the safety-first, ethical lab. So when it's the only holdout — and saying no also happens to protect its paid business — people start asking if it's really about principle, or about the money. Anthropic's own model, Mythos, can find and exploit zero-day bugs on its own. That's the strongest proof yet that AI offense is beating defense. So yes, they're right about the danger. But they built that danger, locked it behind Project Glasswing, and told everyone to trust that closed access stays safe — right before the preview got accessed within a day through an outside vendor. So it's all of it at once: a real argument, convenient timing, and a genuine hit to their reputation. On the business side, though? Nothing's sticking. Claude still tops the benchmarks and enterprises still pay. A company can be isolated and dominant at the same time. For Anthropic right now, it's both principled and protectionist — and they've clearly decided being right matters more than being liked. Amodei's full position: https://lnkd.in/gZP9Ftr8 #AI #Anthropic #OpenSource #OpenWeights #Cybersecurity #AIPolicy #AISafety #Claude #MachineLearning #TechPolicy