The AI Security Podcast

In this episode, Tania and I discuss the debate around closed or open model weights. What do you think?The RAND report we mention: https://www.rand.org/pubs/research_reports/RRA2849-1.html

In this episode, Tania and I talk through some creative examples of prompt injection/engineering we've seen in the wild.. think prompts hidden in papers, red-teaming and web-scraping.Your Brain on ChatGPT: https://arxiv.org/pdf/2506.08872Paper with hidden text (p. 12):  https://arxiv.org/abs/2502.19918v2Interesting overview: https://www.theregister.com/2025/07/07/scholars_try_to_fool_llm_reviewers/Echoleak blog post: https://www.aim.security/lp/aim-labs-echoleak-m365

This week we discussed multiple AI vulnerabilities, including Echolink in M365 Copilot, Agent Smith in Langchain, and a SQL injection flaw in Llama Index, all of which have been patched. We also covered a data exposure bug in Asana's MCP server and OWASP's project to create an AI vulnerability scoring system, while also outlining Google's defense layers for Gemini, Thomas Roccia's Proximity tool for MCP server security, news regarding AI and legal/security concerns, and research on AI hacking AI, prompt compression, multi-agent security protocols, and the security of reasoning models versus LLMs.

I'm back from holiday, and this week Tania and I talk about a project she completed as part of the ARENA AI safety curriculum to replicate the findings of evaluations on frontier AI capabilities.Link to reasoning paper: https://arxiv.org/abs/2502.09696Link to the Inspect dashboard: https://inspect-evals-dashboard.streamlit.app/ARENA AI Safety course: https://www.arena.education/

This week we try a new condensed format for the AI security digest! we covered critical CVEs, including vulnerabilities in AWS MCP, Llama Index, GitHub MCP integration, and tool poisoning attacks. We also reported on malware campaigns using spoofed AI installers, a supply chain attack via fake PyTorch models, and the AI-guided discovery of a Linux kernel vulnerability by Sean Healin using OpenAI's 03 model. We addressed OpenAI's actions against malicious use of their models, Reddit's lawsuit against Anthropic for data scraping, the creation of an AI model for reconstructing 3D faces from DNA by Chinese researchers, a zero-trust framework for AI agent identity management proposed by the Cloud Security Alliance, research on an agent-based red teaming framework, the impact of context length on LLM vulnerability, and CSIRO's technique for improving deep fake detection. We also highlighted the vulnerablemcp.info project and the ongoing evolution of AI security best practices.Sign up to get the digest in your inbox: http://eepurl.com/i7RgRM

Sign up to receive in your inbox: http://eepurl.com/i7RgRMTania Sadhani and Miranda R discussed various AI security topics, including critical CVEs affecting platforms like ChatGPT and Hugging Face, the potential for SharePoint Copilot in internal reconnaissance, and malicious npm packages targeting Cursor developers. They also covered the OASP Gen AI security initiative's Agent Name Service (ANS), the proposed AI.txt for controlling AI agent interactions, and Unit 42's framework for agentic AI attacks. Furthermore, Miranda highlighted security guidance from international agencies, Anthropic triggering ASL 3 for Claude Opus 4, Microsoft's AI red teaming playground, a significant data leak from an AI vendor, and the Israeli police's use of AI-hallucinated laws.

Ever wondered how security vulnerabilities are found in AI? Join us as we chat with Aditya, a Vulnerability Researcher at Mileva Security Labs!

Sign up to receive in your inbox: http://eepurl.com/i7RgRMThis week we note regular CVEs in AI libraries such as Nvidia TensorFlow and PyTorch. We discuss a novel prompt injection technique called "policy puppetry", along with malware dispersal through fake AI video generators and Meta's release of an open-source AI security tool set including Llama Firewall. We also covered Israel's experimental use of AI in warfare, Russia's AI-enabled drones in Ukraine, China's crackdown on AI misuse, Dreadnode's research on AI in red teaming, geolocation doxing via multimodal LLMs, safety research on autonomous vehicle attacks targeting inference time, Config Scan for analyzing malicious configurations on Hugging Face, Spotlight as a physical solution against deepfakes, and Reply Bench for benchmarking autonomous replication of LLM agents.

Alberto Chierici, a physicist-turned-data scientist and author, dives into his remarkable journey through AI and ethics. He examines the balance between human intuition and machine learning, especially in insurance. Alberto discusses the ethical implications of AI technologies, advocating for integrated approaches in engineering. He also explores the evolving tech landscape in Australia and the potential of AI to enhance productivity and well-being. With anecdotes and insights, he paints a hopeful picture for the future of AI.

We talk about Stanford Human-Centred AI's latest AI Index report, check it out here: https://hai.stanford.edu/ai-index/2025-ai-index-report