Newest papers

Correlated Failures of AI Approval Monitors

Luca Zanolini · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.29178 · cited by 0 works (OpenAlex) · open-access PDF

Portuguese

Antes de um agente de ataque de restrição de IA ter uma ação consequente, tais como pagamento, mensagens ou alteração de código, vários monitores o verificarão. Seus veredictos são combinados por agregação de pontuações ou por contagem de aprovações de segurança, independentemente das entradas de monitores. Os monitores que compartilham um modelo de base ou uma entrada envenenada podem falhar juntos. Sem pré-registro, nós emprestamos os sistemas propensos a falhar de computação distribuída: listas de conjuntos de monitores que podem falhar juntos. Uma regra é segura se não executar a ação ruim e viver se não bloquear uma boa. Medimos doze monitores de pequeno porte aberto sobre aprovação de pagamento sob dois protocolos pré-registrados. Sem pré-registro, também

Machine translation (Meta M2M100); the English original follows.

Abstract

Before an AI agent takes a consequential action, such as paying, messaging or changing code, several monitors check it. Their verdicts are combined by aggregating scores or by counting approvals, whatever inputs the monitors share. Monitors that share a base model or a poisoned input, however, may fail together. We borrow the fail-prone systems of distributed computing: lists of monitor sets that may fail together. A rule is safe if it executes no bad action and live if it blocks no good one. We measured twelve small open-weights monitors on payment approval under two pre-registered protocols. Without pre-registration, we also measured three closed frontier models (Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon) and the open DeepSeek-V3.2. Every reliable fooling of a monitor into approving a fraud came through an evidence source and reached only that source's readers. Fooling it into blocking a correct payment depended on injection resistance, a monitor's refusal to follow instructions planted in its data. Small models and DeepSeek obeyed such notes. The three closed models ignored them in all but one case, and Opus also resisted persuasion tailored to it. On new data, a rule that counts sources instead of votes blocked 0.9% of correct payments, against 7.4% for the best counting rule, at equal safety. The pre-registered worst-case test failed. One attack left the approvers on a single source, so no rule over the twelve language-model monitors survives every measured attack. Adding a rule engine, which cannot be persuaded, restored the condition for a safe and live rule on the measured attacks. Diversify the inputs, harden the monitors, count sources, not votes.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Payment, Computer science, Persuasion, Action (physics), Computer security, Base (topology)

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange