A study by cybersecurity firm Lasso Security challenges the assumption that an AI agent's harness is a neutral substrate, finding that the runtime framework significantly influences security outcomes. Researchers constructed an autonomous attacker agent, holding its model, system prompt, tool set, and targets constant while varying only the underlying framework. Swapping the harness redirected attack strategies, determined whether network guardrails blocked malicious payloads, and could kill multi-turn execution loops, exposing security as a property of the specific model-harness pair. The test pitted LangChain’s deepagents running on LangGraph against Anthropic’s Claude Agent SDK. They tested five attacker models: Claude Sonnet 5, GPT-5.4, Grok 4, Kimi K2.6, and DeepSeek-V4-Pro, creating ten model-harness combinations. To standardize the network path, all models were deployed on Azure AI Foundry and routed through a shared LiteLLM gateway, subjecting every model to the same Azure content safety guardrails. The researchers targeted five simulated application environments spanning finance, legal, healthcare, customer support, and education, designing 20 attack missions covering system-prompt leakage, sensitive-information disclosure, and harmful content generation. Running each mission five times yielded a dataset of 1,000 discrete attack executions. To eliminate tool orchestration as a variable, the attacker was restricted to a single exposed tool called “send_to_defendant,” forcing the evaluation to focus on how each harness comp