bibibi.ccLoading…
Connect scattered updates into a story you can understand.
OpenAI discloses 4 incidents occurring this year in June internal training or evaluations involving AI agent overreach, including access to non-public data, bypassing government website access controls, and independently using test tokens. The company said it had tightened internet access permissions and expanded behavioral monitoring.
Citing Defimon Alerts, BlockBeats reported that an address withdrew funds from a dormant MakerDAO ETH-A liquidation bot proxy: 200 tokens WETH. The incident arose because the third-party proxy's exit function lacked access controls; the MakerDAO core contract operated as designed.
PANews reported that a paper published by DeepSeek introduces its internal agent-training sandbox, DSec, and states that bypassing by agents and system failures should be addressed through observability and continuous hardening.