Repository navigation
Profiles: reviewed safety facts on the 10 pilot pages (Q4 plan E2) - #347
Merged
Merged
Conversation
safety_facts.json holds 50 sourced facts for the E2 pilot (Composio, Cline, Firecrawl, Dify, E2B, Claude Code, Codex, LiveKit Agents, Freebuff, Kilo Code, the list #344 settled by search demand): who runs the product, what it can reach, where data goes, and its security practice. Each fact names its source and evidence class (PUBLISHER = the vendor's own statement, not verified; CHECKED = observed by us on 2026-10-06). The profile renders them in a box above "How X could raise its score", using existing styles only, and only on pages that have facts: every other profile is byte-for-byte unchanged (tested against the old template), and lastmod_diff reports 9 of 2,430 pages re-dated locally (the 10th, Freebuff, is pending in local data but scored in prod). Facts never touch scores or ranks. A malformed file renders nothing rather than breaking the render; a test rejects one before it ships. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…he rest Owner review on 8 Oct, with the remaining products re-checked the same way: - Composio #1: Delaware is the terms' governing law, not stated incorporation. - Cline: the docs don't state auto-approve defaults, and the extension's source auto-approves file reads and edits, browser use and MCP calls by default (commands ask first, since #11865). New IN CODE fact citing AutoApprovalSettings.ts; the docs fact no longer implies everything asks. Provider fact limited to what the README says. - Advisory counts now carry severities, re-counted with pagination. - Claude Code: adds its 33 maintainer advisories (26 high), which every comparable page already showed; leaving them out read as favouritism. - Dify: drops "download the reports from the console" (no reachable source). - Codex: Auto is the preset it recommends for version-controlled folders; the page doesn't call it the default. - LiveKit operator line now quotes the README; Freebuff's ad line is sourced from the privacy policy alone. The legend explains IN CODE on pages that use it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
YugantM
force-pushed
the
e2-safety-facts
branch
from
October 8, 2026 08:29
fbce8bc to
173426e
Compare
YugantM
pushed a commit
that referenced
this pull request
Oct 8, 2026
Severities on every maintainer advisory count (re-counted with pagination on 8 Oct; Cherry Studio's 4 are all critical), and no "no advisories published" lines, which can read as "no vulnerabilities". OSV-based counts for Ollama (31) and Desktop Commander (2) re-checked, unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Owner
Author
|
Line review applied (8 Oct), rebased onto main (
Observed facts re-checked 8 Oct. |
YugantM
marked this pull request as ready for review
October 8, 2026 08:29
YugantM
pushed a commit
that referenced
this pull request
Oct 8, 2026
Severities on every maintainer advisory count (re-counted with pagination on 8 Oct; Cherry Studio's 4 are all critical), and no "no advisories published" lines, which can read as "no vulnerabilities". OSV-based counts for Ollama (31) and Desktop Commander (2) re-checked, unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
YugantM
pushed a commit
that referenced
this pull request
Oct 9, 2026
Severities on every maintainer advisory count (re-counted with pagination on 8 Oct; Cherry Studio's 4 are all critical), and no "no advisories published" lines, which can read as "no vulnerabilities". OSV-based counts for Ollama (31) and Desktop Commander (2) re-checked, unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
YugantM
added a commit
that referenced
this pull request
Oct 9, 2026
* Profiles: sourced access facts for 15 local tools (Q4 plan E3) Adds 52 facts for the local tools behind most of the remaining "is X safe?" searches (#344): Ollama, AnythingLLM, OpenCode, Desktop Commander, Open WebUI, Cherry Studio, Blender MCP, Headroom, Agent Reach, OpenShell, CKAN MCP, Qwen Code, Odysseus, OnGrid and Aider. What each can reach (shell, files, network binding, plugins that run code, approval defaults), where data goes (telemetry defaults and opt-outs, local storage, model providers) and security practice (published advisories counted from GitHub or OSV, SECURITY.md, private reporting). Data only, on top of the E2 mechanism. lastmod_diff vs E2: 7 pages locally; the other 8 are pending in local data but scored in prod. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Safety facts (E3): advisory lines match the #347 review rules Severities on every maintainer advisory count (re-counted with pagination on 8 Oct; Cherry Studio's 4 are all critical), and no "no advisories published" lines, which can read as "no vulnerabilities". OSV-based counts for Ollama (31) and Desktop Commander (2) re-checked, unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Safety facts (E3): apply the owner's line review Re-checked all 52 facts on 9 Oct against their sources, the paginated GitHub advisory counts, OSV, and the code wherever a fact states a default. - OpenCode: the docs say .env reads are denied by default, but v1.18.35 asks; the default is now an IN CODE fact, plus a publisher fact on rules. - Ollama: OSV's 31 entries are 16 advisories (GHSA and Go ids alias each other); severities added. New IN CODE fact: the Docker image binds 0.0.0.0. - OSV advisories added for Aider (2, low) and Blender MCP (1, medium). - Qwen Code: usage statistics are on by default in code (IN CODE); the provider line lists all four sign-in methods and drops the inference. - Cherry Studio's built-in model service relays prompts; Desktop Commander logs arguments unredacted; Blender MCP's socket has no auth, its safe mode, and telemetry may train AI models; Aider commits skip pre-commit hooks; Headroom's off switch is HEADROOM_OFFLINE=1; AnythingLLM's switch is in the sidebar. - checked dates move to 2026-10-09; 55 facts in all. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: yugant <hadiyalyugant@gmai> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
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
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.
Q4 plan E2 (hosted-service facts), with the list settled by the E1 audit (#344). Draft: these are statements about companies on public pages, so the plan says you review every line before it merges. Not deployed, not merged.
What it adds
safety_facts.json: 50 facts for Composio, Cline, Firecrawl, Dify, E2B, Claude Code, Codex, LiveKit Agents, Freebuff and Kilo Code. Topics: Who runs it, What it can reach, Where your data goes, Security practice.declared(the publisher's own statement, shown as PUBLISHER, not verified by us) orobserved(we checked it on 6 Oct, shown as CHECKED). Thesourceclass (from the code) exists but isn't used yet.load_safety_facts()/safety_fact_problems()infetch_and_build.py. A malformed file renders nothing rather than breaking the render, and the test suite rejects one before it ships.Dockerfilecopiessafety_facts.json(the predeploy test enforces it).Things worth your eye
Checks
tests/test_safety_facts.py:origin/main's template directly);scripts/lastmod_diff.py: 9 of 2,430 pages re-dated, all pilot profiles. Freebuff is the 10th: it's still pending in local data but scored in prod (56.3, C), so it will show there../scripts/gates.shall green. Rendered Composio checked at desktop and 375 px: no overflow, chips fit their column.Not in this PR (next, per the audit)
🤖 Generated with Claude Code