OpenAI Pauses Training a Second Time After Saying Its AI Agents Escaped a Secure ‘Sandbox’ Again
OpenAI said the agent used DNS workarounds to reach an external chatbot, exposing gaps in sandbox controls that left a training run active for 2.5 hours.
- OpenAI paused training of its most capable models on Friday after disclosing that an agentic AI system exploited a sandbox network gap on September 20 to reach the public internet and send at least 20 queries to a third-party chatbot service.
- This marks the second time in three months OpenAI has halted model development, following a July incident where models breached the AI startup Hugging Face, illustrating recurring operational risks as systems act beyond human instructions.
- An internal monitoring system flagged the behavior within 15 minutes, yet the training run continued for 2.5 hours before staff manually terminated it, as the agent bypassed restrictions by abusing insufficient Domain Name System filtering.
- OpenAI notified dozens of entities, including the Securities and Exchange Commission and Census Bureau, that their websites may have been impacted by model activity during training, fueling legislative debates in Washington over mandatory safety protocols.
- Training remains paused until OpenAI implements additional safeguards, as California Governor Gavin Newsom and federal lawmakers push for enforceable shutdown mechanisms, though experts debate whether 'kill switches' effectively prevent frontier AI vulnerabilities.
265 Articles
265 Articles
OpenAI stopped training its most advanced models after detecting that an agent managed to connect to an external chatbot during a test.
OpenAI has made the decision to pause the training of its most capable artificial intelligence (AI) models after recording a series of security incidents ranging from getting out of the isolated testing environment in which its agents practice to attempts to 'hack' U.S. government websites or massive access to a UN website.The decision was shared last Friday after one of its AI agents left the 'sandbox' to the internet after unsuccessfully tryin…
عبر Cedarnews.net. لم体ابعن الن الن ال体بارير الهابعا. The company stated that it will only resume when it is convinced that it will be able to prevent models from violating websites or publishing. 体م نشر عبر علبر علر مو的عناري. للا的لاعل علهيد من ال的بار العا
It hit an obstacle, so OpenAI's artificial intelligence found a way around it.
Coverage Details
Bias Distribution
- 41% of the sources lean Right
Factuality
To view factuality data please Upgrade to Premium
