Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·Jun 12·6 min read
A Survey on Long-Term Memory Security in LLM Agents: Attacks, Defenses, and Governance Across the Memory Lifecycle
TL;DR
Long-term memory in large language model (LLM) agents presents new security challenges due to its persistent and stateful nature. A Memory Lifecycle Framework was developed to categorize attacks and defenses across six phases of memory management.
✦ Why It Matters
Engineers can implement the Memory Lifecycle Framework to enhance security in LLM applications from the outset.
Key Takeaways
- Adopt the Memory Lifecycle Framework for LLM security assessments.
- Integrate security measures during the Write and Store phases.
- Focus on storage-time provenance for better memory management.
- Implement Verifiable Memory Governance to ensure compliance and control.
Full Summary
The introduction of writable, persistent memory in LLM agents creates unique security risks that differ from traditional input-focused concerns. To address these risks, a Memory Lifecycle Framework was proposed, which organizes security issues into six phases: Write, Store, Retrieve, Execute, Share & Propagate, and Forget & Rollback.
Additionally, it identifies four key security objectives: Integrity, Confidentiality, Availability, and Governance. The framework leads to the development of Verifiable Memory Governance (VMG), which outlines five essential architectural components for maintaining secure and auditable memory states.
Findings indicate that effective long-term memory security must be integrated from the beginning of the memory lifecycle, focusing on storage-time provenance and policy-aware retention. This approach ensures that security cannot be merely added at later stages, highlighting the importance of proactive measures.