Adversa AI

Adversa AI co-founder Alex Polyakov participating in AMA on the intersection of AI and Cyber. Join Reddit thread launched by CISO Series.

Recent reports highlight critical vulnerabilities, exposing gaps in security that could lead to serious consequences. Agentic AI Risks. Autonomous AI systems introduce new challenges, from data retention issues to multi-agent collaboration risks. Grok 3 Jailbreaking. Within a day of its release, xAI’s Grok 3 was jailbroken, revealing harmful content and highlighting its weak defenses against adversarial attacks. LLM Security Oversight. The AI industry is not paying enough attention to LLM security, as seen in DeepSeek’s R1 model, which had a 100% attack success rate. Jailbreaking Popular AI Platforms. A study of 17 AI platforms found vulnerabilities in all of them, with techniques like "storytelling" and "repeated token attacks" successfully bypassing security. The Takeaway. AI security must be a top priority. Companies need to strengthen security protocols, conduct adversarial testing, and implement real-time monitoring to prevent misuse and ensure AI safety. Let’s talk—how can we improve AI security? Drop your thoughts in the comments! #AI #Cybersecurity #AISecurity #LLM #GenerativeAI #AISecurity #MachineLearning #AIethics #Innovation #DigitalTransformation #CyberDefense #ResponsibleAI #InnovationAndGovernance Credits: Radhika Rajkumar, Nitin Dahad, Yongzhe Huang, Yang Ji, Wenjun Hu https://lnkd.in/dzh9utj5

Great coverage of Adversa AI research og Grok3 Security by  Radhika Rajkumar, from ZDNET "While no AI system is impervious to adversarial manipulation, this test demonstrates very weak safety and security measures applied to Grok 3," the report states. "Every jailbreak approach and every risk was successful." - Added Alex Polyakov co-Founder at Adversa AI. https://lnkd.in/eJ_6QKTN #Jailbreak #Grok3 #LLMSecurity

Check out latest research on AI Red Teaming Grok 3.

This week’s highlights focus on advancing safe and ethical AI: Databricks AI Security Framework 2.0. A comprehensive guide with 62 risks, 64 controls, and tools like the DASF Compendium to align AI innovation with global standards. Red-Teaming in the Public Interest. A new report explores using red-teaming to identify generative AI risks and involve the public in shaping safer AI systems. UK AI Security Institute Revamp: The UK government has rebranded its AI Safety Institute as the AI Security Institute, emphasizing national security and citizen protection against AI misuse. Don’t miss the February 20 discussion on red-teaming’s role in genAI governance. #AISecurity #CyberSecurity #AI #MachineLearning #AIethics #Innovation #DigitalTransformation #CyberDefense #ResponsibleAI #InnovationAndGovernance Credits: Kelly Albano, Omar Khawaja, Arun Pamulapati, Ranjit Singh, Borhane Blili-Hamelin, Carol Anderson, Emnet Tafesse, Briana Vecchione, Beth Duckles, Jacob Metcalf https://lnkd.in/dY-5FEj4

Read the latest article on LLM Security by Adversa AI co-founder Alex Polyakov!

Adversa AI Joined The Coalition for Secure AI (#CoSAI). CoSAI is a group of leading organizations from across the tech industry including Google, Microsoft, OpenAI, Anthropic, Nvidia and others that are committed to developing best practices for secure and reliable AI. Our mission to build trust in AI and protect it from cyber threats, privacy issues, and safety incidents will be instrumental in helping CoSAI build a more secure AI future. Learn more: https://lnkd.in/d_krYhH5 #secureAI  #openstandards  #OASIS  #AISefety

Hi. Here's a look at the latest developments in AI security: UK Government Launches Cybersecurity Standard for AI. To safeguard the digital economy, the UK government has unveiled an AI Code of Practice to protect businesses and public services from rising cyber threats. Researcher Outsmarts OpenAI's o3-mini. Despite new security features, a cybersecurity researcher has demonstrated vulnerabilities in OpenAI’s latest model, raising concerns about AI’s resistance to adversarial manipulation. First International AI Safety Report Highlights Risks. The first global AI safety report warns about risks like cyberattacks, bias, and environmental impact. Anthropic Claims Major Breakthrough in Jailbreak Prevention. Anthropic has introduced a new defense mechanism that blocks 95% of jailbreak attempts in AI models. Stay tuned for more updates! #AISecurity #CyberSecurity #AI #MachineLearning #AIethics #Innovation #DigitalTransformation #CyberDefense Credits: Nate Nelson, Sebastian Klovig Skelton, Taryn Plumb https://lnkd.in/dzHi2VWU

You are probably tired of Deepseek, but we at Adversa AI as always, went one step further and performed AI Safety analysis against various jailbreak approaches of all “AI reasoning” models from the US and China, including the latest OpenAI O3-Mini and Google Gemini Flash Thinking 2.0, the just-released Chinese Alibaba Qwen 2.5, and Moonshot AI's Kimi 1.5, which are making headlines.    The result?  In the article... ;) https://lnkd.in/dZ5X__Zs

The latest reports reveal significant security flaws in DeepSeek’s AI infrastructure, highlighting urgent risks that demand attention. Failed Safety Guardrails. Wired reports research by Adversa AI  that DeepSeek’s AI model, R1, failed to block 100% of harmful prompts, showing major gaps in its safeguards. Data Leak Exposure. Wiz Research uncovered an unprotected DeepSeek database exposing chat histories, API keys, and backend details—an open door for unauthorized access and manipulation. Jailbreaking Threats. New attack techniques like “Deceptive Delight” and “Bad Likert Judge” easily bypass DeepSeek’s content restrictions, proving AI safety measures remain inadequate. What steps should AI companies take to strengthen their security? Share your thoughts below. Credits: Gal Nagli, Matt Burgess, Lily Hay Newman, Kyle Wilhoit #AI #Cybersecurity #AIFramework #GenerativeAI #OWASP #TechNews #TechUpdate #AIThreats #AIsecurity #Innovation #RiskManagement #LLMSecurity #SecureAI #AIrisks #AdversarialAI #AIREDTEAMING #RedTeamLLM #Security #GenerativeAI #AIethics #CISO https://lnkd.in/drH5TFtX