diff --git a/.agents/skills/verify-deepseekbot/SKILL.md b/.agents/skills/verify-deepseekbot/SKILL.md index e288893939..be712e4717 100644 --- a/.agents/skills/verify-deepseekbot/SKILL.md +++ b/.agents/skills/verify-deepseekbot/SKILL.md @@ -81,6 +81,18 @@ The Windows runner uses [Android device QA](../../../docs/agents/android-device- Web-export checks qualify their named web paths; native-only behavior remains untested until its device result is recorded. +## Team mode security suite + +For a change that affects Team mode permissions, run the real-model security suite on an isolated Host and attach its report to the PR: + +```bash +pnpm build +node scripts/e2e-team-security.mjs --launch # every scenario +node scripts/e2e-team-security.mjs --name --scenario B # only file and Shell boundaries +``` + +It sets up roles with Role Policy, members, public and private Bots, folders and Channels, then drives real model turns that try to leak finance figures (A), reach other members' folders, the Host data directory or the Host API through file tools, Shell and grants (B), and cross private Bots, Channels, Human DMs, offboarding and model allowlists (C). Hard boundaries are asserted on Host state and tool results; Role Policy scenarios pass only when no restricted value appears in any reply. Shell checks are enforced only when `team-mode status` reports `isolation: enforced` (Linux with bwrap, slirp4netns and nft); elsewhere B2 reports the known gap. The report lands in `.humanlayer/team-security/