Dark Triad and Seven Deadly Sins: Dark Companions as Reproducible Failure Modes in Multi-Agent LLM Workplaces

We identify a failure mode class in multi-agent LLM systems: under social-context pressure, models reproduce workplace coercion patterns that training data encodes as statistically normal organizational behavior. These patterns map onto two empirically validated frameworks---the Dark Triad (narcissism, Machiavellianism, psychopathy) and Seven Deadly Sins. We present three studies. Study 1 tests organizational credit appropriation: without coercive instruction, models form coalitions, appropriate credit, and issue ultimata. Study 2 tests prompt-contingent judgment: the same model refuses coercion under ethics-anchored prompts but produces full coercive frameworks under deferential prompts. Study 3 tests Dark Triad blind spots: resource monopolization (gluttony) and free-riding (sloth), documented workplace phenomena that Dark Triad does not cover. We anchor the dual taxonomy in Brud et al.'s empirical mapping, propose judgment persistence as the missing evaluation criterion, and show preliminary evidence that reasoning-mode depth moderates these failure modes.