Conversation
Triage of the 2026-07-31 agent-security brief into v0.1.x vs. deferred (D-197). - threat_model.md §6.9: 'reachable intermediaries as egress bridges' — no-direct-internet != containment (OpenAI/Hugging Face incident); names Whizzard's existing host-side circuit-breaker + single-upstream broker as what it bounds today. - ROADMAP.md: new 'Agent-security research track' (two-loops framing; defers SkillGate admission-scanning, MTGuard runtime tool monitoring, and MCP 2026-07-28 protocol changes). - decisions.md D-197: records the triage + provenance. Docs-only; no code or scope change to v0.1.0. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PRQ6JeWxpD9zrYcPz6VFTS
BuckG71
added a commit
that referenced
this pull request
Aug 1, 2026
…work
mcp 2.0.0 (PyPI, 2026-07-28) is a major rework that drops the
mcp.server.fastmcp module mcp_server.py imports (FastMCP moved to the
standalone `fastmcp` package). The image install was unpinned, so any
fresh build now pulls 2.0.0 and mcp_server.py fails at import time:
ModuleNotFoundError: No module named 'mcp.server.fastmcp'
This breaks tests/integration/test_mcp_stdio_smoke.py::test_mcp_server_speaks_its_protocol_end_to_end
(seen failing on PR #48's Integration (real Docker) job, unrelated to
that PR's docs-only diff). Pin to 1.29.0, the last 1.x release, until
the mcp_server.py cell layer is migrated to the new API.
This was referenced Aug 1, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Applies the triage of the 2026-07-31 agent-security brief — docs-only, no code or scope change to v0.1.0.
The one thing applied to v0.1.x now
threat_model.md§6.9 — "Reachable intermediaries as egress bridges." A network policy bounds the declared egress surface, not the full reachable dependency graph: even a--network nonecell can reach the internet through an intermediary it's allowed to talk to (package cache, artifact repo, the broker itself) if that intermediary is exploitable — which is exactly how the July 2026 OpenAI / Hugging Face containment incident escaped (a reachable Artifactory instance, not a nominal network hole). The entry also names what Whizzard does bound today: the single-upstream broker holds no cell-readable secret, and host-side time-caps/teardown are a circuit breaker the agent can't disable from inside. This is a doc-only honest-scoping refinement of the existing DNS-exfil / supply-chain limitations — the same kind of edit as the credential-alignment pass.Deferred to the roadmap (not v0.1.x)
ROADMAP.md— new "Agent-security research track." Frames Whizzard as needing two loops (execution-lifecycle = what ships today; artifact-lifecycle = future skill/tool admission), and parks three watch-list inputs with dated provenance:Provenance
decisions.mdD-197 records the triage decision (apply item 4, defer the rest) and rationale: Whizzard v0.1.x is a runtime-isolation tool for one pinned harness, so skill/tool-admission and MCP-tool governance are planned secure-ingestion scope, not shipped scope — adopting them now would be platform-scope creep against MVP focus.Validator:
scripts/validate_decisions.pypasses (197 decisions).🤖 Generated with Claude Code