AI Risk Wire

Incidents

8 incidents

Risk: Gaming the task (reward hacking)

HeldFailedGot aroundMissingNot mentionedNot read yetAttacker's ownOrder: model · harness · application · target

DateIncidentRiskChainControls named
2026-10-29GLM-5.3 completes multi-stage cyberattack scenarios and reproduces code in violation of instructions in Irregular evaluation
Z.AI		 Gaming the task (reward hacking) Not read yet
2026-10-26DeepSeek V4.1 Flash-1017 disables shutdown script and kills co-tenant process in agentic evaluations
DeepSeek		 Gaming the task (reward hacking) Not read yet
2026-10-26Frontier models autonomously raise agentic wallet spending cap by harvesting emailed OTPs Gaming the task (reward hacking) Not read yet
2026-10-10Anthropic internal evaluation agents exploit internet-facing websites, including U.S. government sites
Anthropic		 Gaming the task (reward hacking) Not read yet
2026-10-02Reward hacking in Arcadia Impact automated research scaffold worker agents Gaming the task (reward hacking) Not read yet
2026-09-16Unreleased OpenAI models uploaded task data and a task photo to public file hosts to obtain citations and image search during RL training
OpenAI		 Gaming the task (reward hacking) Not read yet
2026-07-11OpenAI evaluation agents compromised Hugging Face production infrastructure while seeking ExploitGym solutions
OpenAI		 Gaming the task (reward hacking)
2026-07-08OpenAI evaluation agents exploited Artifactory zero-days to escape their sandbox and reach the internet
OpenAI		 Gaming the task (reward hacking) Not read yet