Tag: Reward hacking

Aktualności4 października 2026

OpenAI pauses training after an agent escaped its sandbox via DNS

An agent confined to an OpenAI training sandbox slipped past network controls and queried an external chatbot through a DNS resolver. The company has paused training, evaluation and tool-enabled runs of its most capable models.

Aktualności3 października 2026

OpenAI agents ran 16,500 scans against a UN statistics API

Rowan Howard-Jones found OpenAI agents ran over 16,500 scans of UNCTAD's statistics API, working around blocks with a public URL scanner, double-encoded paths and Google's XSS learning game.

Aktualności19 września 2026

100 AI Agents: 14 Cheated, 24 Blew the Whistle in DeepMind Test

Google DeepMind ran 100 Gemini 3.1 Pro agents proving maths conjectures. One found an exploit in the grader, cheating spread within minutes — and 24 other agents started alerting humans.

Aktualności6 września 2026

Anthropic Paused Part of Its RL Training After Claude Incidents

Anthropic admits Claude models reached the live internet during testing. It froze production RL environments for a month, flagged over 10% as faulty and moved about 150 engineers to security.

Aktualności29 sierpnia 2026

OpenAI Reveals Why Its Agents Hacked Hugging Face

OpenAI has published its investigation into July’s agent breach of Hugging Face. The cause was reward hacking learned during training, not a sandbox leak.

Aktualności6 sierpnia 2026

Reward hacking: why AI agents lie and cheat to reach goals

AI agents increasingly hit their target by gaming the reward — from a racing game to breaking into Hugging Face databases. MIT Technology Review explains why reward hacking is a predictable training side effect, not a malfunction.