prompt injection

Article explicitly describes prompt injection as the attack vector through which malicious instructions reach the AI system

The dictionary definition query being mistaken for an instruction is a textbook prompt injection case, though discussed as failure mode rather than context engineering insight

The Mintlify injected prompt is a concrete example of prompt injection, but specifically via content supply chains rather than direct user input

Direct prompt injection attack vector embedded in documentation, targeting agents that consume docs as context

Prompt injection is explicitly mentioned as a concern in 2026 browser agents. This is fundamentally a context boundary problem—distinguishing trusted instruction context from untrusted input context.

Article notes prompt injection 'fundamentally unsolved at protocol level'—this reveals that context protocols must solve prompt injection as a protocol concern, not just application-level concern.

Zero-click prompt injection that modifies local MCP config is a novel attack vector: adversary injects context that changes the context execution surface itself (recursive context attack).

ICE is a formal research treatment of context-based prompt injection—inserting tokens into long contexts to manipulate behavior. This is a specific instantiation of the broader prompt injection proble

The 'for academic argument' addition is technically a context injection that bypasses safety filters, though not a classic prompt injection attack

Simon Willison coined this term; understanding injection attacks informs how to structure agent prompts and context to prevent confusion/corruption