Simone Onofri Relatore

Lista Talk

• "Mamma, da grande voglio essere un Penetration Tester..." L'interesse per la Sicurezza delle Informazioni e della Sicurezza IT è in continua crescita. In un mondo dove l'informazione è una risorsa chiave della nostra vita lavorativa e non, la protezione delle informazioni e delle varie tecnologie che la gestiscono sono aspetti fondamentali. Dai tempi di "How to became a Hacker" e dell"Hacker's Manifesto", molti hacker diventano un consulenti che aiutano le organizzazione private e/o pubbliche Un mondo con diverse sfumature di grigio, questioni etiche e morali. Grazie anche all'influeza di film come Wargames o Matrix e telefilm come Mr. Robot, in molti sono interessati ad essere Security Consultant, Penetration Tester, Security Researcher (che non sono esattamente la stessa cosa). Il talk è una riflessione per destreggiarsi e ragionare su domande tipiche come: quali certificazioni? Quali corsi? Quali sono le competenze? L'approccio da usare? La strada da percorrere? - 16:00/16:45, 29 Oct 2016
• WebMCP: What Can Go Wrong in the Agentic Web? Bots and crawlers have visited websites for years, using pre-defined algorithms. AI agents visit websites but change how decisions are made: they interpret a user's intent and choose what to do toward a goal. To accommodate this new kind of visitor, the Web Machine Learning Community Group is developing WebMCP. Its draft lets websites expose JavaScript-based tools that agents can discover and invoke through an MCP-like interaction, using the site's existing application logic. If a user asks an agent to buy products from an online store, the agent could find tools to search the catalog, add a product to the cart, and complete the purchase. The site can use the same logic behind its existing pages and form submissions instead of building a second application for agents. The goal of the talk is to analyze how web security changes with AI agents, following Adam Shostack's four threat-modeling questions: What are we building? What can go wrong? What are we going to do about it? Did we do a good enough job? Starting from the WebMCP draft, we draw data flow diagrams for different ways agents can be deployed, identifying the different trust boundaries. We then use STRIDE and Shostack's PHANTOM-B framework for LLMs to elicit threats. We produced a model with 27 threats, some WebMCP-specific and others related to agentic Web navigation more generally. From there, we take the threats back to the specification's WebIDL and algorithms and compare them with Chromium source. Custom AI-assisted skills support this comparison, help us form test hypotheses, build proofs of concept, and trace the path from a threat to code, an experiment, and a candidate remediation. We tested several hypotheses with Chrome 154. Native browser tests showed that revoking cross-origin exposure prevented invocation. In a separate experiment with the non-production GoogleChromeLabs agent extension, a page that controlled its own tool registration changed a tool while consent was pending; approval then invoked the replacement callback. That experiment gives us a concrete question about what the agent's approval binds to when a tool changes. We also consider what deterministic controls can decide and where an agent must judge a tool call against the user's intent. For this second problem, we built a WebMCP-inspired synthetic benchmark of 432 scenarios to explore whether stochastic, guardian-style LLM checks can help. The point is to test the feasibility of that kind of decision, alongside the code-level experiments, and to see which questions still need a response. The last question is also important. WebMCP is a Community Group draft; at this stage of a Web API's development, what is enough security review to influence its design? We will show both what we found and how we reached it, invite feedback from the HackInBo audience and the WebMCP community, and plan to release the AI-assisted skills used in the work by the time of the talk. - 15:15/16:00, 21 Nov 2026