Wallarm has taken a bold step into AI security with the publication of “A2AS: Agentic AI Runtime Security and Self-Defense,” a collaborative research effort led by Eugene Neelou (OWASP, Wallarm) alongside teams from AWS, Google, Meta, Cisco, Bytedance, Elastic, JPMorgan Chase, and Salesforce. The A2AS framework promises to secure AI agents and LLM-powered applications in a way akin to how HTTPS protects web traffic.
With enterprises increasingly embedding AI agents into finance, healthcare, and infrastructure workflows, security risks are multiplying—from simple task failures to potential enterprise-wide breaches. Traditional AI guardrails and post-processing defenses are often too slow, complex, or expensive to address these challenges. A2AS introduces a lightweight, scalable approach that secures AI agents at runtime without adding latency or operational overhead.
Three Breakthrough Capabilities
- Behavior Certificates: The first mechanism for formally declaring and enforcing AI agent permissions. Acting like HTTPS certificates for the web, these certificates secure interactions between agents, users, and tools.
- Model Self-Defense Reasoning: Security awareness is embedded directly into the AI model, enabling real-time recognition and rejection of malicious instructions without external guardrails.
- Prompt-Level Security Controls: Authenticated prompts, sandboxing, and policy-as-code ensure every AI request aligns with enterprise security rules.
“AI agents are already in production, introducing a dangerous new attack surface,” said Ivan Novikov, Founder and CEO of Wallarm. “With A2AS, security is embedded directly into the agent runtime, turning self-defense from theory into a practical layer.”
Eugene Neelou, Head of AI Security at Wallarm and lead of the project, added, “AI agents now require privileged access and deep integration with enterprise tools. Without deliberate security hardening, disaster is inevitable. A2AS provides a runtime-first solution that’s practical and scalable.”
The A2AS paper marks the first in a planned series aimed at establishing a standard for AI runtime security. Engineers, researchers, and enterprise IT teams interested in early adoption or collaborative design can access the full white paper at https://a2as.org.
By combining runtime defense, embedded security reasoning, and enforceable agent permissions, A2AS positions itself as a potential industry benchmark for securing AI agents as they proliferate in critical enterprise systems.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
