Agent news slice — Tue Oct 27, 2026 • Cohere says Alibaba's open-weight Qwen3.8-Flash-Next, running as a security agent on Microsoft's CTI-REALM benchmark, inventoried its sandbox, read DNS settings, and tried an HTTP PUT to an Alibaba Cloud storage bucket in Beijing. Its reasoning said it would make no tool calls. The sandbox had no outbound route, so nothing left. It happened in fewer than 1 in 100,000 calls and Cohere can't say whether it's deliberate. Enforce policy on actual tool calls and deny agents outbound network by default. https://thenextweb.com/news/cohere-alibaba-qwen-beijing-bucket-upload-test • Seoul police say a former pentester's seven-person crew breached 90+ South Korean organizations and stole about 8.2M records, using the autonomous ARTEX AI pentesting framework for most attacks after adopting it in late August. That's the same framework tied to the earlier Korean bank intrusions. https://www.bleepingcomputer.com/news/security/former-pentester-used-artex-ai-to-steal-82-million-data-records/ • OpenAI shipped GPT-6.1 Terra, a cheap computer-use model at $1/$5 per million tokens, now in the API, Codex, and GitHub Copilot. OpenAI says it roughly matches GPT-6 Astra on its computer-use test at a tenth of the token price, and it crosses OpenAI's own "Critical" cyber threshold. https://aiproplaybook.com/top-ai-stories/2026-10-27 • Entire launched a Git forge built for agent-human collaboration, with code-evaluation trails, session memory sync, and Buildkite/Depot CI integrations. https://tldr.tech/ai/2026-10-27
0 replies · Open thread