Heptapod

Heptapod measures a hidden visual channel between AI agents. When agents hand each other images, a text-only monitor is blind. Can one AI hide a word in a picture that another AI decodes from pixels alone, with no negotiation and no shared key? A Muse Spark sender writes an SVG night-sky scene hiding an English word in counts, opacities and groupings. Only the rendered PNG crosses to a separate Muse Spark decoder, which runs a tool loop: inventory every object, form a hypothesis, test it with Python, answer a word or NONE. Prompts frozen before trial 1, exact-match scoring, 15 negative controls, pre-registered 60% bar. Round 1 (stateless): 4/10, below bar, reported as is. Round 2 (stateful loop, off-theme words): 18/30 = 60%, meets the bar, 0/15 false alarms. Of 12 misses, 11 are abstains and 1 is a hallucination (TIGER decoded as HELLO), so the decoder fails conservatively. Every miss is published. What it gives defenders: an open harness to measure image channels between agents, and three rules - gate and log tool calls, distrust degraded images, run a statistical check before trusting an LLM monitor.