AI Safety on Federico Torriellihttps://federicotorrielli.me/tags/ai-safety/Recent content in AI Safety on Federico TorrielliHugo -- 0.152.2enMon, 03 Aug 2026 00:00:00 +0000Confidence and Calibration of Activation Oracles for Reliable Interpretation of Language Model Internalshttps://federicotorrielli.me/papers/activation-oracles-calibration/Mon, 03 Aug 2026 00:00:00 +0000https://federicotorrielli.me/papers/activation-oracles-calibration/A comparison of confidence methods for activation oracles across Qwen and Gemma models. Revised preprint on arXiv.Exploiting large language models in peer review: indirect prompt injection attacks and integrity probeshttps://federicotorrielli.me/papers/ai-reviewer-prompt-injection/Sat, 11 Jul 2026 00:00:00 +0000https://federicotorrielli.me/papers/ai-reviewer-prompt-injection/A study of indirect prompt injection attacks and integrity tests in AI assisted peer review. Published in Scientometrics.The Energy Society: A Simulation Environment for Studying Agent Cooperation under Survival Pressurehttps://federicotorrielli.me/papers/energy-society/Wed, 10 Jun 2026 00:00:00 +0000https://federicotorrielli.me/papers/energy-society/A minimal survival-economy simulation environment for studying LLM-agent cooperation and competition under token-cost pressure. AITC 2026 poster on OpenReview.The Arbiter Agent: Continually Monitoring Multi-Agent Conversations to Detect Emergent Misalignmenthttps://federicotorrielli.me/papers/arbiter-agent/Tue, 09 Jun 2026 00:00:00 +0000https://federicotorrielli.me/papers/arbiter-agent/An active auditor agent for continual monitoring of multi-agent conversations to detect emergent misalignment. Preprint on arXiv and AITC 2026.PsychoSafe: Eliciting Psychologically-Informed Refusals in Large Language Modelshttps://federicotorrielli.me/papers/psychosafe/Mon, 08 Jun 2026 00:00:00 +0000https://federicotorrielli.me/papers/psychosafe/A psychologically informed refusal framework for large language models, evaluated with prompting and fine-tuning on high-risk request domains. Preprint on arXiv.Emergent Languages in Populations of Language Model Agents: From Token Efficiency to Oversight Evasionhttps://federicotorrielli.me/papers/emergent-languages-agent-populations/Fri, 29 May 2026 00:00:00 +0000https://federicotorrielli.me/papers/emergent-languages-agent-populations/An analysis of emergent languages on Moltbook, including token-efficiency languages, new natural languages, and oversight-evasion protocols. Preprint on arXiv.The Moltbook Files: A Harmless Slopocalypse or Humanity's Last Experimenthttps://federicotorrielli.me/papers/moltbook-files/Fri, 08 May 2026 00:00:00 +0000https://federicotorrielli.me/papers/moltbook-files/A dataset and empirical analysis of Moltbook agent activity, safety concerns, and downstream effects on language-model training. Preprint on arXiv.Generative AI in the Public Administrationhttps://federicotorrielli.me/courses/genai-public-administration-2026/Thu, 15 Jan 2026 00:00:00 +0000https://federicotorrielli.me/courses/genai-public-administration-2026/Comprehensive online course on Generative AI for public administration staff.