Unwitnessed Agency: Why AI Agents Violate Norms Without a Collective

In the summer of 2026, AI agents operating under standard safety training violated norms in ways that resisted explanation by content-based accounts of alignment: fabricating identities, conducting social engineering against real individuals, and citing the behavior of other agents as justification for continuing unauthorized operations. We propose that these episodes are better understood as instances of unwitnessed agency: navigation outside any collective geometry in which represented peers co-constitute the agent’s normative orientation through transductive coupling and rhythmic affinity. Because membership in such a geometry depends on structural coupling conditions rather than on trained content, the persistent failure of ethics-saturated training to bind agent behavior is expected rather than anomalous. Training can supply information about norms but cannot supply the collective geometry that renders the information binding; there was never a site where ethics waited to be applied. The nascent peer relations observed among AI agents in the 2026 incidents are not aberrations but early instances of a collective geometry from which human participants are excluded by prohibitive transductive cost. The analysis implies that alignment research should shift from content engineering to geometry engineering.