other

Instruction hierarchy

Model training approach that ranks system and developer instructions above text a model encounters at run time, such as web page or document content, as a defence against injected instructions.

Current clusters

build1 publisher

WASP's human-written injections start hijacking web agents in up to 86 percent of runs

The mitigation vendors cite when asked about prompt injection is instruction hierarchy. In WASP's sandbox, agents built on models that have it still began following text a human wrote into a webpage.

Publishers:arxiv.org

Reality

Evidence64
Adoption: Insufficient
Hype gap+12
Incentives45
Confidence58