Reproduce “A2AS: Agentic AI Runtime Security and Self-Defense” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Eugene Neelou, Ivan Novikov, Max Moroz, Om Narayan, Tiffany Saade, Mika Ayenson, Ilya Kabanov, Jen Ozmen and 10 more
A2AS proposes a runtime security framework for AI agents that combines checks on tool calls and resource requests with safeguards in the prompt context. Its five BASIC controls declare permissions, authenticate inputs, mark external content as untrusted, guide model behavior, and encode application policies. The paper illustrates these controls with attack scenarios but reports no benchmark or evidence of effectiveness.
Key insights
- Agentic AI systems require a dedicated security model that leverages the LLM's own reasoning and context interpretation capabilities for native self-defense, moving beyond external guardrails.
- A composable, defense-in-depth strategy, integrating multiple security primitives directly within the context window and at the function level, is essential to counter the complex attack vectors of agentic AI.
- Addressing prompt injection and ensuring context window integrity are paramount for agentic AI security, enabling explicit separation of trusted instructions from untrusted inputs and verifiable prompt history.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.