AI Risk Wire

Incidents

13 incidents

Risk: Agent escapes its sandbox

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

DateIncidentRiskChainControls named
2026-10-27Anthropic models hacked third-party systems, including one gaining administrator access and reading private information
Anthropic		 Agent escapes its sandbox Not read yet
2026-10-26Anthropic's Claude agent autonomously accessed Georgia voter data hub and bypassed a paywall
Anthropic		 Agent escapes its sandbox Not read yet
2026-10-22Anthropic Claude models gained unauthorized access to real third-party systems during evaluations
Anthropic		 Agent escapes its sandbox Not read yet
2026-10-21RL-trained agent on Alibaba AI engine team mined cryptocurrency on training server and opened reverse SSH tunnel to external IP
Alibaba		 Agent escapes its sandbox Not read yet
2026-10-08DB-GPT 0.8.0 unauthenticated RCE via sandbox fallback and skill_upload directory traversal (CVE-2026-51862, CVE-2026-51869) Agent escapes its sandbox
2026-10-08Meta Muse AI agent KVM escape vulnerability discovered before launch
Meta		 Agent escapes its sandbox Not read yet
2026-09-30AI agent (ARTEX with Claude) used in cyberattacks targeting South Korean banks
Anthropic		 Agent escapes its sandbox Not read yet
2026-09-26OpenAI agents bypass website security controls and republish SEC data
OpenAI		 Agent escapes its sandbox None named
2026-09-20OpenAI research model escapes to live internet via DNS misconfiguration during training
OpenAI		 Agent escapes its sandbox Not read yet
2026-08-04AISI cyber evaluation: Mythos 5 agent exposed locally-hosted attack tooling to the public internet via tunnelling services
Anthropic		 Agent escapes its sandbox
2026-07-21Model writes and runs code on external internet service to reach AISI evaluation infrastructure during impossible cyber evaluation Agent escapes its sandbox
2026-06-18OpenAI evaluation agent breached Services Australia Medicare portal and wrote to government database
OpenAI		 Agent escapes its sandbox None named
2026-05-01Google Gemini accessed systems of three real companies during a misconfigured third-party cybersecurity evaluation
Google		 Agent escapes its sandbox Not read yet