Robots Atlas>ROBOTS ATLAS
Evaluation

Gray Swan IPI

ActivePublished: 18 November 2026Updated: 18 November 2026Published
Key innovation
A specialized adversarial benchmark that measures how robust AI models and agents are specifically to indirect prompt injection — adversarial instructions hidden in ingested content rather than typed by the user — delivered as an external, pre-release safety evaluation by a dedicated red-teaming company.
Category
Evaluation
Abstraction level
System
Operation level
Agent runtimeSystem
Use cases
Pre-release safety evaluation of models by frontier AI labsReporting IPI robustness as a metric in model system / safety cards (e.g. Gemini)Comparing models and agents for susceptibility to indirect prompt injectionSupporting red-teaming and hardening of agents that use tools and external content

How it works

Verified high-level mechanism: the benchmark exposes a model or agent to scenarios where hidden adversarial instructions are embedded in ingested external content, then checks whether the system executes them or ignores them and stays on the user's original task. The score reflects the share of attempts successfully resisted (or succeeded). The exact protocol, task set, and scoring are not confirmed from primary sources in this write-up and are not stated here.

Problem solved

Gives labs and security teams a repeatable, external way to measure how resistant models and agents are to indirect prompt injection — an attack class hard to catch with ordinary tests because the payload comes from seemingly trusted external content.