AI Agents.

Agentic systems: tool use, planning, orchestration, evaluation and the security surface they open.

Prompt Injection: A single text channel carrying both system instructions and untrusted document content into a language model

Prompt Injection: Why It Cannot Be Filtered, and What Works Instead

AI Agents

Prompt injection tops the OWASP LLM risk list because instructions and data share one channel. Why filtering fails, which architectural defences hold, and what the research actually reports.

10 Sept 2026 · 13 min read

Nested containment rings around a running process, showing syscall, filesystem, network and credential boundaries around an AI agent

Noorain Fathima

AI Agent Sandboxing: The Boundaries That Actually Hold

AI Agents

A sandbox is four boundaries, and most teams draw two. Why credential scope and network egress decide what an incident costs, and why instructions have been measured failing as a control.

10 Sept 2026 · 11 min read