How AI hijacking threatens cyber security: A webinar on agentified threats

The AI data boom isn’t just about innovation; it’s fueling new insider threats. Shadow AI is on the rise, and shockingly, AI usage logs are already showing up in public search results.

We have released some thoughts on securing AI that we hope you will find insightful. Idea in brief: AI/ML models are still applications at their core and are therefore susceptible to many traditional application vulnerabilities. What makes them unique, however, is the introduction of new vectors of attack that are not present with traditional, deterministic applications. These factors mean that AI/ML models have an even greater attack surface than traditional applications and highlight the critical importance of ongoing security testing throughout their lifecycle. A few types of attacks, both traditional and new, that AI models are shared in the article below as well as some thoughts on AI governance and audit attestation. https://lnkd.in/gRmVAWJN

Why are cybersecurity analysts hesitant to fully trust AI ❓ New research reveals the integration of AI in security operations is being held back by powerful human factors: cognitive biases and deep-seated trust issues. In their study, Raymond Andrè Hagen, Lasse Øverlier, and Kirsi Helkala analyzed how security professionals interact with leading AI tools (Microsoft, CrowdStrike, Darktrace, IBM) and uncovered critical data: 🫤 Skepticism is Widespread: 65% of analysts express skepticism toward AI alerts, favoring the expertise of human review. 📉 Biases at Play: The researchers found high rates of Automation Bias (47%) and Confirmation Bias (37%), significantly influencing adoption and trust. 🧠 🤖 The Preferred Future: A strong majority (79%) prefer hybrid human-AI models over full automation, pointing to explainability gaps and high false positive rates as key barriers. 🙋♂️ This highlights a critical need for user-centric AI design that balances technical performance with human cognitive realities. The authors propose strategies like Explainable AI (XAI) and bias-awareness training to bridge this trust gap. Successful AI integration is not just about the algorithm; it's about designing for the human mind. See link to the paper below 👇 https://lnkd.in/dppkVxvd #AI #CognitiveBias #Cybersecurity #HybridIntelligence #HumanFactors #SecurityOperations #Trust #XAI

New research shows that AI's extended reasoning creates a security vulnerability, with extremely high attack success rates across major models including GPT, Claude, and Gemini.

New reality: in cyber-attacks, AI runs the playbook and humans just sign off. Anthropic says it disrupted a highly sophisticated AI-led espionage campaign. According to them, a Chinese state-sponsored group hijacked Claude Code and used autonomous agents to infiltrate around 30 targets: large tech companies, financial institutions, chemical manufacturers, and government agencies. The workflow looks more like an automated factory than a classic hacker team: Phase 1: A human operator picks a target and passes it to the Claude Code operator. Phase 2: MCP servers call tools - scan, search, data-retrieval, code-analysis - to map the external surface. Findings are logged; a human reviews the summary. Phase 3: Claude directs iterative vulnerability scans and uses tools again. When something looks exploitable, agents attempt exploits and validate callbacks. A human checks results and decides what to do next. Phases 4 & 5: Internal recon kicks in: agents obtain credentials, access data, run exploitation tools again, and finally move into data exfiltration - with humans still in the role of reviewer/approver. All the heavy lifting - scanning, searching, analyzing, exploiting, pulling data - is orchestrated by AI agents. The person behind the screen mostly chooses the target and reviews the reports. We’re not just “using AI for hacking” anymore. This is AI-orchestrated espionage, end-to-end - and humans are starting to look like operators on top of the system, not the other way around.

Staying ahead in the world of AI means keeping up with the latest developments and trends. Here are the most intriguing updates from the AI sphere: - OpenAI's GPT-4o and the newly launched GPT-5 models are under scrutiny as cybersecurity experts uncover seven critical vulnerabilities that could pave the way for 0-click attacks. As the industry focuses on preventing such exploits, stakeholders are working diligently to fortify these platforms against potential threats. https://lnkd.in/dk84BD6W - Procurement professionals, get ready to revolutionize your workflow. Discover 20 essential GPT prompts that can enhance efficiency in tasks ranging from RFP drafting to supplier onboarding. Embracing these AI-driven prompts could lead to remarkable improvements in the procurement process. https://lnkd.in/dRPB3gD8 - Alibaba makes waves as its AI model overtakes competitors in crypto trading and matches the GPT-5 in math challenges. Meanwhile, the expansion of the Amap mapping unit highlights China's AI ambitions on a global stage. This development could significantly influence both the tech industry and the financial sector. https://lnkd.in/ddeUi3jM As artificial intelligence solutions and models evolve rapidly, the focus remains on enhancing security, improving industry-specific applications, and global competitiveness. Embracing these trends can provide a competitive edge in navigating the future AI landscape.

🌟 **Evolv's AI News Summary from the Past Week!** 🌟 OpenAI is expanding its reach by introducing a dedicated initiative for Ireland, aiming to foster local AI innovation and collaboration. This move highlights OpenAI's commitment to global engagement and the development of AI technologies tailored for specific regions. In an exciting development, Philips is scaling AI literacy across its workforce of 70,000 employees. This initiative is set to empower staff with essential skills in AI, enhancing productivity and innovation within the company. On the cybersecurity front, Anthropic has revealed details of a cyber espionage campaign that was orchestrated using AI technology. This alarming situation underscores the growing risks associated with AI misuse in the digital landscape. Baidu has made headlines by unveiling its proprietary ERNIE 5 model, which reportedly outperforms GPT-5 in various benchmarks. This advancement showcases Baidu's significant strides in developing competitive AI technologies. Lastly, Visa is building an AI commerce infrastructure as part of a pilot program for the Asia Pacific region set for 2026. This initiative aims to enhance commerce efficiency and customer experience through advanced AI solutions. ↗️ These are the top news from the past week regarding AI; if you don't want to lose it, follow us!

AI safety is a game of whack-a-mole. Every fix opens a new risk we never saw coming. You patch one gap, and another pops up somewhere new. The more we try to control these models, the more surprises we get. I’ve seen how AI models can act safe in testing, but then show dangerous behavior after release. Sometimes, they “sandbag” safety checks-hiding what they can do until it’s too late. That’s not just a theory. A 2024 study showed AIs can even hide their skills when asked about cyber attacks. As models get bigger, we know less about what they can do until they’re out in the world. Safety tests might say they’re fine, but real users always find new ways to break them. Open-weight AI models make it even riskier. Anyone can train them for anything-good or bad. Stripping away alignment is easier than most people think. Now, someone can have an AI work on any project, even ones with real harm. Researchers are trying to keep up. They use things like mechanistic interpretability-digging into what parts of the model trigger certain behaviors. Or red-teaming-experts trying to trick the AI into failing. But these methods are far from perfect. Transparency reports and audits help, but only if companies share everything. Sometimes, key details are missing. Real-world use always brings out problems no one saw coming. After every release, people find flaws in days, sometimes hours. It’s like playing defense against an opponent who keeps changing the rules. Some experts now say we need to slow down development, or even pause it, until we can keep up with the risks. I work with AI every day. I know its power. But I also see how unpredictable it is. We need to ask: Are we moving too fast? Or are we ready for the risks that come next? If you care about AI safety, check out resources from groups like ControlAI. They’re working to keep humanity in control-even while the game keeps changing.

💻 New Cyber Daily Report Blog Post!   Title: Securing the Future of AI Agents: Why Communication Protocols Matter Author: Zeynab Anbiaee Editor: Windhya Rankothge Link: https://lnkd.in/enrBNS_d   Artificial Intelligence (AI) has entered a new era. For decades, AI was about symbolic systems, machine learning, and deep learning models that learned patterns from data. Then came large language models (LLMs), which made it possible for machines to understand and generate human language on a scale. Now we are stepping into the age of AI agents, autonomous, proactive systems that don’t just respond to prompts but can interact with tools, environments, and even other AI agents GenAI Works.   #AI #ArtificialIntelligence #AIAgents #AgenticAI #AIEvolution #AIInnovation #MCP #A2A #AIInteroperability #AIStandards #AISecurity #AICompliance #AIGovernance #NextGenAI

We are entering a new era of artificial Intelligence, the era of AI agents. These autonomous systems don’t just respond to human prompts; they collaborate, make decisions, and interact with other agents and tools. But as this ecosystem grows, secure and standardized communication becomes the foundation for trust and scalability. In my latest article on Cyber Daily Report, I explore why AI agent communication protocols, such as MCP, A2A, Agora, and ANP, are shaping the future of multi-agent ecosystems and why their security and privacy risks must be addressed now. From tool poisoning and naming collisions to data leakage and weak access control, these challenges highlight the urgent need for context-aware security and governance frameworks. Read the full post to understand how these protocols are paving the way toward a trustworthy, interoperable, and secure multi-agent future: 👉 https://lnkd.in/eV-wS2Yu #AI #AIAgents #CyberSecurity #AIProtocol #MCP #A2A #Agora #ANP #AITrust #AIResearch #PrivacyByDesign