An attack surface is the set of points through which an attacker can attempt to affect or compromise a system. In a generative AI workflow, it includes not only surrounding software but also the model and the content it reads. Data access, write access, and the ability to act expand what an attacker may be able to influence. section 4.0 section 1.1
A language model interprets instructions and reference material in natural language rather than enforcing a reliable technical boundary between them. An attacker can exploit that ambiguity by placing instructions in content the assistant is supposed to process. Filters can stop obvious attempts, but they are not an enforced boundary. section 4.0 section 4.6.3
A poisoned retrieval path
An attacker who can place a document or email where an assistant searches may hide instructions in it. If the assistant retrieves that content, the instructions can influence its response or actions. The combination of private-data access, untrusted content, and an outward communication path can enable exfiltration. section 4.6.3
Match controls to the exposed capabilities. Narrow search scope and exclude or de-prioritize loosely governed sources; sanitize documents before indexing to remove hidden formatting and metadata; and require source citations so unexpected inputs are visible. Policy engines, tiered approval, and sandboxing can constrain actions as capabilities expand. These measures reduce exposure and influence rather than guarantee that retrieved instructions will be ignored. section 4.6.3 section 4.0






