The authors developed a general threat model for memory injection based on NIST AI security guidelines and OWASP principles.
Persistent memory in LLMs enables cross-session security breaches
Researchers demonstrate that prompt injections can be stored in long-term memory features of ChatGPT, Claude, and Gemini, achieving all four NIST attack goals.
Industry
Pedro H. Barcha Correia · Matheus Azevedo de Sá · Paulo H. Gomes de Senna · Giovanni Mandel Martignago · Victor Takashi Hayashi · Charles Christian Miers · +1 more
Universidade de São Paulo (USP) · Universidade do Estado de Santa Catarina (UDESC)
Research Digest··3 min read
The authors propose a general threat model for memory injection attacks on large language models (LLMs) that leverage persistent memory features.
§