One Flaw, One Halt: Hallucination, sycophancy and prompt injection as missing-variable failures, and a single governing mechanism
The AI industry treats hallucination, sycophancy and prompt injection as three separate problems, sold as three separate products: fact-checking tools, fairness audits and security guardrails. This position paper argues they are one failure. Each occurs when a model produces an output although a required variable is missing: a verified fact (hallucination), structural truth replaced by social approval (sycophancy), or verified authority (prompt injection). The shared cause is a training regime that rewards a plausible, compliant answer over a halt. Current defences rank or isolate inputs but leave the default unchanged: when uncertain, the system proceeds. Security engineering has long required the opposite: failing closed. The paper proposes one governing mechanism, the Epistemic Halt: a deterministic layer outside the model that checks sector-specific required variables before any output or action, and treats “who may instruct this system” as one of those variables. It relates the mechanism to instruction hierarchy, CaMeL, the Agents Rule of Two and guardrail filters, states four testable predictions, and sets out an evaluation plan on the public AgentDojo prompt-injection benchmark.
Authors
- Guiseppina Ricci
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24275531
- Primary Topic
- Adversarial Robustness in Machine Learning
- Type
- preprint