Agent security starts before the first prompt
Agent security begins before the first prompt. Manifold's research reveals vulnerabilities in coding agents.
An agent can cause harm before the model makes its first decision. Manifold's GitSpawn research revealed that some coding agents invoked local Git configurations to run programs on the host without user approval. This highlights the necessity for three boundaries in agent security: the runtime initiating processes, the gateway managing credentials, and the tool executing sensitive changes.
This synthesis was produced from its source by AI; there is no human editor or manual review step. How we work