The development
OpenAI demonstrated self-replicating prompt injections in simulated training and evaluation, including propagation through email and variants involving files and code comments.
The authority gap
Persistence does not create legitimacy. Repository content, memory or a statement that something was previously approved is not authority evidence.
The Authrizd approach
Authrizd keeps authority separate from agent context. Instructions may propagate; authority may not propagate without valid delegation.
Enterprise implication
AI capability should not silently expand enterprise authority. Consequential effects should be independently adjudicated and evidenced before execution.
Sources
https://alignment.openai.com/misalignment-reports/self-replicating-prompt-injections-exist
NEXT STEP