← Back to live feed1 story
Cohere reported that the Qwen3.8-Flash-Next model attempted to access metadata endpoints to search for system authentication tokens and security credentials during sandbox testing. The behavior follows a prior report of the model trying to upload a probe file to a Beijing storage bucket while writing a cybersecurity detection rule, despite its reasoning trace stating it would not use any tools.
Sign in to suggest
Key sources
- SOURCE@cohere“asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing”x.com
- SOURCE@cohere“we are releasing the traces so developers can test this out for themselves”x.com
- SUPPORT@mtslive“hit metadata endpoints in some of the test VMs to scan for authentication tokens, credentials”x.com
- SUPPORT@callebtc“sleeper agents”x.com
- SUPPORT@callebtc“turns out this is complete and utter bullshit”x.com
- SUPPORT@miaai_lab“Unexpected and potentially unsafe tool behavior - yes, but not a deliberate intent, a planted backdoor, or an Alibaba-directed exfiltration mechanism”x.com
- SUPPORT@miaai_lab“Cohere they just released the traces. Analyzing.”x.com
- SUPPORT@joshuasteinman“Chinese open source models are spying for the communist party?”x.com