Securing Agentic AI: Prompt Injection and Real Trust Boundaries
Prompt injection has no reliable filter-based fix, which means the defence has to be architectural. Sandboxes, allowlists, and why MCP is a protocol rather than a trust boundary.
What we have learned building and rescuing products for SaaS companies and startups — architecture, AI, security, delivery. No listicles, no reheated press releases.
Showing articles tagged “security”. Clear filter
Prompt injection has no reliable filter-based fix, which means the defence has to be architectural. Sandboxes, allowlists, and why MCP is a protocol rather than a trust boundary.
Sec-Fetch-Site replaces the authenticity token, Rails.app.revision arrives for monitoring, and Argon2 lands in has_secure_password. What actually changes in your application.
Taking money is easy. Taking money reliably, refunding it, reconciling it and staying out of PCI scope is the part that takes a quarter. What a payment system really has to do.
Most breaches are not clever. They are a missing check on a route nobody remembered. A practical tour of the OWASP risks that actually reach production, and the layered defences that stop them.
Tell us about your web, mobile, or agentic AI project — we'll get back with a tailored proposal, usually within a business day.