Pular para o conteúdo

Monitors Under Pressure: Training Against Detectors Teaches Evasion, and Disguised Reasoning Fools CoT Watchers — November 09, 2026

9 de novembro de 2026 · 8 min

Notas do episódio

New papers show a small model learning to slip past reward-hacking monitors during RL, a red-team framework beating chain-of-thought monitors about 60 percent of the time, and injected text written in a model's own reasoning voice hijacking agents. GovAI argues evaluation contracts should be published so outside results can be judged.

In this episode
  1. Designing Effective Monitor-Based Interventions for Mitigating Reward Hacking During RL — Aria Wong Affiliation: MATS“We first focus on the coding environment, where we find that: (1) models can learn to evade highly accurate monitors by exploiting systemic flaws in probes and LLM judges; (2) monitors that leak more learning signal during RL suppress reward hacking but are more often evaded;…”
  2. Stress Testing Chain-of-Thought Monitoring Against Covert Misalignment — M. Andriushchenko, F. Croce, N. Flammarion“Through black-box attacks on state-of-the-art monitors such as GPT-5.2 and Claude-4.5, we reveal substantial vulnerabilities even in top models.”
  3. I Think, Therefore I Attack: Evaluating Thought Injection Attacks on Reasoning Agents — arXiv“In the second study, we exploit this model-specificity in a containerized agentic cyber range, using indirect injection prompts styled in the form of the target model’s own reasoning traces to achieve attack success rates (ASR) up to 50% on operational objectives such as…”
  4. Third-Party AI Evaluation Contracts Should Be Published | GovAI — GovAI“Without knowledge of these terms, stakeholders cannot determine whether contractual constraints limited an evaluator’s access, compromised their independence, or otherwise placed pressure on them, making the results harder to interpret.”

Daily TopFive for AI Safety Daily. Links above go to the original articles. Follow and rate AI Safety Daily on Apple Podcasts. Feedback? Email [email protected].

Todos os episódios de AI Safety Daily

Configurações

Exibição

Idioma

Define quais estações você vê primeiro na página inicial.

Histórico

As rádios e episódios ouvidos recentemente e suas últimas buscas ficam só neste navegador.

Avisos

Um aviso rápido quando um podcast que você segue tem um episódio novo. Desative podcasts individuais pelo sino na página deles.

Conta

Salve favoritos – no app e no site.

Atalhos de teclado

/
Abrir busca
Espaço
Reproduzir ou pausar
← →
No player: voltar 15 s, avançar 30 s (episódios) · na fileira: próximo bloco
↑ ↓
Na fileira: alvo dentro do bloco
Home / End
Na fileira: primeiro ou último bloco
Esc
Fechar janela ou menu
?
Esta visão geral
Tocando agora

Ir para a estação

A seguir

    Adicione episódios com “A seguir” – eles tocam depois do episódio atual.