A Safety Memory Is a Declassification Channel: Why Origin Binding, Taint Tracking and Memory Isolation Do Not Secure the Records an Agent Writes About Attacks

APA

Mahendrakar, P. (2026). A Safety Memory Is a Declassification Channel: Why Origin Binding, Taint Tracking and Memory Isolation Do Not Secure the Records an Agent Writes About Attacks [Journal article]. Life of Research. https://doi.org/10.5281/zenodo.22262653

Copy APA

BibTeX

@misc{lor32026,
  title        = {A Safety Memory Is a Declassification Channel: Why Origin Binding, Taint Tracking and Memory Isolation Do Not Secure the Records an Agent Writes About Attacks},
  author       = {Pranay Mahendrakar},
  year         = {2026},
  publisher    = {Life of Research},
  doi          = {10.5281/zenodo.22262653},
  url          = {https://lifeofresearch.com/articles/3},
  keywords     = {agent memory, memory poisoning, information-flow control, declassification, integrity, LLM agent safety, guardrails, prompt injection},
  note         = {LOR-00003}
}

Copy BibTeX More on the record page