GitHub Copilot sandboxes agents; GPT-6, Opus 5.5, Grok 4.7 join plans
Fail-closed sandboxing lands as the agent-breach story escalates; VS Code 1.139 runs agents in remote containers, and 950 Claude agents crack a new enzyme.
Copy markdown
Four frontier models, sorted by plan tier
GPT-6 Sol and Claude Opus 5.5 are now selectable in Copilot on Pro+, Max, Business and Enterprise; GPT-6 Luna and Grok 4.7 reach all the way down to Pro. If you've been paying per token elsewhere, the top-tier models are now inside a subscription you may already hold — check your plan before wiring one into an agent loop.
Fail-closed sandboxing for your coding agents
The Copilot app can now box an agent away from your files, network and credentials by default (public preview), with OpenTelemetry traces of everything it touched; JetBrains adds assisted approvals that auto-clear only low-risk tool calls. It's the concrete answer to a week of agent-breach headlines — least privilege for the thing running commands in your repo.
VS Code 1.139 runs agents inside remote Dev Containers
The new release lets Copilot agents run in Dev Containers over SSH, Tunnels and WSL, so the agent works in the same reproducible box your code builds in — plus faster agent sessions and the end of the auto-appended 'generated with Copilot' disclaimer on commits.
950 agents, 21 hours, one new enzyme
Anthropic set roughly 950 Claude agents loose on a genomic database and they surfaced a novel CRISPR-adjacent enzyme family (ARTs), since confirmed in the wet lab — 210M tokens over 21 hours, driven by Claude Code and Claude Science. The takeaway for builders isn't the biology; it's the replicable pattern: point an agent swarm at a dataset too big to read and let it find the needle.