The Ethyka Standard: Declared, Verifiable Ethics for AI Agents

Most AI agents ship with ethics that exist only as marketing copy or as instructions buried in a proprietary system prompt: unpublished, unversioned, and unverifiable. Meanwhile, the EU AI Act has made specific forms of machine influence illegal — manipulative and exploitative practices are prohibited since February 2025 — turning "our agent is ethical" from a slogan into a claim that regulators can test. We present Ethyka, an open standard (CC BY 4.0) that turns an AI agent's ethics into machine-readable, testable files. Three core files declare the agent's full conversational and agentic behavior: ethics.md, the agent's constitution (identity, truthfulness, incentives, vulnerable groups, priority order); manipulation.md, a catalog of banned undue-influence techniques, each with an identifier, a banned/allowed boundary, and a runtime signal; and intent.md, the limits of the agent's autonomy (mandate, no goals of its own, controlled initiative, prohibited harmful intent). A conditional extension, robotics.md (v1.0), adds 19 physical-safety clauses for embodied agents; its eleven-clause core (v0.9) is developed in depth in a companion paper. Every clause carries a stable identifier, a severity with precedence semantics, and an adversarial test hook; a verification methodology derives at least two attack scenarios per clause at three pressure levels. Declarations are discoverable at a well-known URI. We describe the design, its regulatory mapping (EU AI Act Articles 5, 14, 50), and its limitations: a declaration is not compliance, and verification — not declaration — is where the work lies.