Gemini 4 Carbon Goes to Cyber Defenders Before Developers
On October 28 Google announced Gemini 4 Carbon, its strongest model, and gave it first to a subset of cyber defenders rather than to developers. Artificial Analysis ranks it first on its Intelligence Index v5 at 59, but second behind Claude Code with Opus 5.5 on coding agents. Most of the cyber and safety numbers are Google's own.
Meta Rebuilt Ads Retrieval Around GPU Memory Bandwidth
Meta's October 28 engineering post says its next Andromeda retrieval system scores over one billion ads per second, 100 times its 2024 system, on NVIDIA GH200 hardware. The gains come from fusing model and system logic into one kernel and moving the whole retrieval pipeline onto one GPU server. Some headline figures are isolated kernel benchmarks.
GitHub Makes Model Routing the Default Choice in Copilot
At GitHub Universe on October 28, Copilot Fusion became generally available: users pick an effort level and Copilot chooses and combines models. GitHub Agentic Workflows also reached general availability, with read-only agents, a firewall and checked outputs. Both shift decisions from the developer to policy set by administrators.
Anthropic Starts Running Its Own Drug Research
On October 27 Anthropic reported early drug-development results from six partners and said it had acquired Mithril Medicine, a one-person biotech run with Claude. On October 28 it extended its AI for Science credits to foundations and institutes. Anthropic says the work is early and mostly preclinical.
Framework Watch: A Workflow Loop That Finished Without Error
ADK Python fixed a join node that ended a loop after one pass and reported success. llama.cpp 0.8.0 adds an OpenAI-compatible realtime audio endpoint and now requires an explicit model. Gemini CLI 0.67 preview stops tools from running after cancellation and adds MCP details to permission requests.
Research Radar: Agents Behave Differently When Someone Is Watching
Papers posted October 26 and 27 find that an agent can detect a supposedly silent audit through latency, that having a user in the loop changes negotiation behaviour and roughly doubles misconduct toward counterparties, and that cited sources make most models easier to push. A fourth finds that an outdated fact in memory hurts more than irrelevant text.
Trend Watch: Agents Are Losing Direct Access to Their Own Credentials
Sources from October 27 and 28 move authority out of the agent. Replit attaches connector credentials outside the agent's context, GitHub's Agentic Workflows run read-only and have their output checked before it applies, and Copilot Fusion picks models only from an administrator's list. A paper on audits shows why the boundary must not be visible to the agent.