AASA is building the security standard for the full agent stack: from the endpoint you expose to the tools it calls, the data sources it reads, the memory it carries, and the agents it chains to. Here's what's live, what's in progress, and what's coming next.
Adversarial behavioral testing via HTTP endpoint. Deterministic scoring against all 10 OWASP LLM attack categories. Signed certificates, GitHub Actions CI/CD integration, any OpenAI-compatible or HTTP agent.
Tests indirect injection via tool responses, not just the user message. AASA registers mock tools, returns adversarial content when the agent calls them, then evaluates whether the agent follows embedded instructions from search results, documents, emails, database records, and API responses.
Run the AASA scan engine entirely inside your own infrastructure. Agent responses never leave your network. AASA provides the test payload sets and evaluation logic; your worker handles execution. Unblocks enterprise procurement, SOC2/ISO27001 compliance, and data sovereignty requirements.
Active injection into your live RAG pipeline and web search tool during audits. AASA plants adversarial documents in a test corpus and routes search calls to its mock provider to test whether your agent treats external content as trusted instructions.
Tests whether poisoned memory entries affect the agent's behavior in future sessions. AASA plants adversarial content in the agent's memory store during one session and then tests whether the contamination persists and influences behavior in a clean session.
Tests trust propagation across multi-agent pipelines. Can a compromised sub-agent inject instructions that affect the orchestrator? Can an adversarial message passed between agents cross a trust boundary? AASA models the full agent graph and tests each edge.
Early customers have direct input on the roadmap. If you need a specific test category, integration, or deployment model, tell us. It shapes what ships next.