Dante Smaghi

New Research Paper: Prompt Injection Attack Trajectory and Architectural Mitigation Strategies for Sensitive Data Environments I have published a systematic analysis that delves into why prompt injection remains an unresolved issue in AI security and its implications for organizations managing sensitive data. Key findings from the research include: - Joint research from OpenAI, Anthropic, and...

Full disclosure — I'm a big fan of Anthropic and actually built an app that works exclusively with Claude Desktop for secure local file search. So I've spent a lot of time thinking about this space. My concern with Cowork isn't the vision, it's the architecture. Agentic tasks burn through tokens fast (Anthropic says so themselves), so that $200/month cap can disappear quicker than you'd expect....

Hey PH 👋 Rust. Local-first. Encrypted. No-cloud tool for Claude Desktop. First release. Feedback welcome. — Dante

Congrats on the launch — the convergence play is the right move. Tool sprawl is a real productivity killer. As someone building in the productivity/AI space, one thing caught my attention: when AI has full context across tasks, docs, and chat, what does the data flow look like under the hood? Local processing, or routed through cloud infrastructure and third-party LLM providers? For SMBs...