Thanks for the thorough review — all 8 questions are addressed below, and
v1.2.15 (Pro) +
v1.4.36 (Admin) are both live on XenForo RM now with the security fix you flagged.
1. The 5 "missing" bundles — intentional split (v1.2.12, Sept 22)
29 tools moved from Pro into a dedicated add-on,
AI Connect Moderation for XenForo. The tools that moved exactly match your list:
- Thread moderation (13):
editThread, lockThread, unlockThread, stickThread, unstickThread, moveThread, deleteThread, featureThread, unfeatureThread, changeThreadType, markSolution, deletePost, mergePosts
- Profile moderation (2):
deleteProfilePost, deleteProfilePostComment
- Reports (5):
listReports, getReport, resolveReport, rejectReport, replyToReport
- Approval (3):
listApprovalQueue, approveContent, rejectContent
- Warnings (4):
issueWarning, deleteWarning, listWarnings, listWarningDefinitions
- Advanced (1):
splitPosts
- Deletion log (1):
listDeletionLog
AI Connect Moderation v1.0.2 adds 10 more tools (moderation queue + spam intelligence), bringing it to
39 total. If you were using Pro's moderation tools in 1.2.11, install AI Connect Moderation alongside Pro — same tool names, same bundle keys, no agent code change required.
Rationale: moderation permissions sit with a different staff group in most installs. Separating them lets you license and permission them independently from Pro's content-creation surface.
It should be noted that existing customers of the Pro add-on who are affected by this change will receive a free license for the Moderation add-on, ensuring they can continue using these tools. We will be in touch with these customers shortly.
2. Tool count — 84 is correct
Our website correctly lists
84 tools across 14 bundles. The 81/11 in
addon.json was stale metadata left from v1.1.3, before three bundles were added in v1.1.4 (Users & Groups, Watched & Bookmarks, Forum Watch). The real manifest, permissions, and bundle maps all carry the 84/14 figure — only the description string was out of date. Updated in v1.2.15.
3. Core version
Pro 1.2.15 requires
chgold/AIConnect 1.1.9 or later. Your 1.2.57 is fully supported — no breaking changes since 1.1.9. Latest Core is 1.2.65; not required for Pro 1.2.15.
4. Upgrade preservation
Confirmed, all four preserved:
- XF permissions — existing per-group assignments untouched. New permissions are registered with default = "unset" (not accessible) for every group except super_admin, so newly added tools are NOT accessible until you explicitly grant them in ACP → Groups & Permissions.
- OAuth tokens — stored in
xf_aiconnect_token, not touched.
- Licence key — stored as an XF option, not touched.
- Per-tool restrictions — XF content-permission rows are not touched by
postUpgrade.
5. Per-account isolation — three enforcement layers
- OAuth scope: each token carries read/write/delete/admin scopes; tools without the matching scope return
403 scope_missing.
- XF permissions: every tool calls
canView/canEdit/canDelete on the acting visitor. If the user cannot do it via the web interface, the agent cannot do it through the API. Example: Pro's listAlerts uses findAlertsForUser(visitor->user_id), so a Sales agent cannot see a Marketing agent's alerts.
- Per-tool permission: admin allows/denies each of the 84 tools per group, bundled under 14 "AI Connect Pro — {bundle}" sets.
For finer control, goldnat.ai lets you scope each connected agent to a subset of the tools already exposed to it via XF permissions.
6. XFRM / XFMG / third-party
Not supported in Pro itself. The 14 bundles cover native XF content types only (threads, posts, polls, users, usergroups, conversations, attachments, bookmarks, notifications, drafts, member stats, forum watch, analytics).
Expanding coverage to additional XenForo content types is actively on our roadmap. Those capabilities will
NOT be added to Pro itself — they will ship as separate add-ons alongside Pro (same model as AI Connect Admin and AI Connect Moderation). This keeps each add-on scoped to one concern, lets you license and permission them independently, and avoids forcing customers to pay for integrations they do not need. Priority follows customer demand — if XFRM in particular is a specific need for you, let me know.
7. URL attachment uploading — fixed in v1.2.15 + v1.4.36 (both live now)
You were right to flag this. The previous
fetchUpload() in v1.2.14 used PHP
file_get_contents() directly, with no built-in protection against internal network addresses, unsafe redirects, oversized downloads, or slow-loris timeouts.
v1.2.15 switches
ProMediaTrait::fetchUpload (
uploadAttachment +
uploadMyAvatar) to XF's built-in
XF\Http\Reader::getUntrusted() — the same hardened fetcher XF uses internally for Gravatar fetch, image proxy, and oEmbed. It provides:
- Private/reserved IP rejection (127.0.0.0/8, 169.254/16, RFC1918)
- Private TLD rejection
- Scheme/port whitelist (http/https on 80/443 only)
- Per-hop redirect validation (
allow_redirects=false + manual recheck)
- Configurable time (30s) and size limits (50 MB)
AI Connect Admin 1.4.36 (also live now) applies the same switch to its three URL-fetching tools:
uploadUserAvatar,
uploadSiteLogo,
uploadStyleAsset. Admin's style asset tool keeps its 5 MB cap for logos.
Using XF's hardened Reader rather than rolling our own means we inherit any future XF SSRF improvements automatically. Verified end-to-end.
8. icon.png hash in hashes.json
v1.2.15 and 1.4.36 both carry regenerated
hashes.json files for good measure (54 and 51 entries respectively, all matching).
Thanks again — this is exactly the kind of feedback that catches real issues before they bite other customers.
Security — defense-in-depth on agent-writable config
What's new
