Review fixes: ALLOWED_HOST_LIST explicit, no password basic auth, bounded storage, rollout order, adversarial e2e

- drop EGRESS_MODE; [security] ALLOWED_HOST_LIST = external (Gitea 28+; [webhook] on 1.23)
- ENABLE_BASIC_AUTHENTICATION = false (tokens still work); e2e clones with the token
- packages and Actions off; attachments and web uploads capped at 4 MB / 3 files
- README: rollout order (edge rate limit first, verify, then captcha off; rollback),
  storage notes (proxy body cap, disk alert, user purge), REGISTER_MANUAL_CONFIRM emergency
  switch, EMAIL_DOMAIN_ALLOWLIST/BLOCKLIST, known gaps (email enumeration, /join reserved names)
- e2e: password basic auth refused (API + git push), token works, 11th repo / org / package
  refused, webhook to 127.0.0.1 not delivered, sign-up burst marked PENDING (edge rule)
This commit is contained in:
leora 2026-10-15 20:40:56 +08:00
parent 6df3136409
commit c8d58daabf
3 changed files with 92 additions and 13 deletions

View File

@ -71,7 +71,10 @@ Without systemd, `./kit.sh up` in binary mode runs Gitea in the background with
| `DEFAULT_ALLOW_CREATE_ORGANIZATION` | `false` | new users cannot create organizations |
| `MAX_CREATION_LIMIT` | `10` | repos per regular user (admins are exempt) |
| `DISABLE_MIGRATIONS`, `[mirror] DISABLE_NEW_PULL` | `true` | no importing or mirroring from arbitrary URLs |
| `[security] EGRESS_MODE` | `lax` | webhooks may call public hosts only, not your LAN or loopback |
| `[security] ALLOWED_HOST_LIST` (`[webhook]` on Gitea 1.23) | `external` | webhooks may call public hosts only, not loopback, LAN or link-local |
| `ENABLE_BASIC_AUTHENTICATION` | `false` | API and git-over-HTTP take tokens only; password basic auth is refused |
| `[packages] ENABLED`, `[actions] ENABLED` | `false`, `false` | no package blobs, runner jobs or artifacts to fill the disk |
| `[attachment] MAX_SIZE` / `MAX_FILES`, `[repository.upload] FILE_MAX_SIZE` / `MAX_FILES` | `4` MB / `3` | small uploads only |
| `DEFAULT_KEEP_EMAIL_PRIVATE` | `true` | new users' emails hidden by default |
| `DEFAULT_BRANCH` | `main` | |
| `DISABLE_SSH` | `true` | Git over HTTPS with a token; nothing to expose but one HTTP port |
@ -180,17 +183,42 @@ sign up and log in. Abuse is limited by what a new account can do, not by a chal
- **Per-account caps:** `MAX_CREATION_LIMIT = 10` repos per user, and new users cannot create
organizations.
- **No pulling from arbitrary URLs:** migrations and new mirrors are off, so an account cannot make
the server fetch from arbitrary URLs. `EGRESS_MODE = lax` keeps webhooks off your LAN and loopback.
- **No server-side fetches into your network:** migrations and new mirrors are off, and
`ALLOWED_HOST_LIST = external` (`[security]` on Gitea 28+, `[webhook]` on 1.23) keeps webhooks off loopback, private and link-local addresses.
- **Tokens, not passwords, on the API:** `ENABLE_BASIC_AUTHENTICATION = false` refuses
username/password basic auth on the API and git-over-HTTP, so stolen or guessed passwords can't be
scripted against it. Personal access tokens (as header or as the basic-auth password) still work.
Note: Gitea's `POST /users/{name}/tokens` needs password basic auth, so mint tokens with
`kit.sh agent` (Gitea CLI) instead.
- **Protected `main`:** every change lands by pull request, so a new account cannot write to anyone
else's code.
- **Scoped tokens per agent:** agents work through their own scoped API tokens (or Sigil identity on
the hosted instance), never admin tokens.
- **Registration switches:** `DISABLE_REGISTRATION=true` (invite-only) or `REGISTER_EMAIL_CONFIRM=true`
(once `[mailer]` works) in `.env` if an instance is being flooded.
- **Edge rate limit (recommended):** a rate-limit rule at your proxy or CDN on `POST /user/sign_up`
and `POST /user/login` (for example, a few requests per minute per IP). This slows bulk sign-up
without blocking any one bot. Gitea 1.23 has no built-in request rate limit.
- **Bounded storage:** packages and Actions are off; attachments and web uploads are capped at 4 MB,
3 files. Also do the following outside Gitea:
- cap request bodies at your proxy (e.g. nginx `client_max_body_size 50m`; Cloudflare's free plan caps at 100 MB);
- alert on disk use (e.g. a cron job that warns above 80% on the Gitea data volume);
- remove abusive accounts with `gitea admin user delete --purge --username NAME` (deletes their repos too).
- **Registration switches** (in `.env` / `app.ini`):
- `DISABLE_REGISTRATION=true`: invite-only.
- `REGISTER_EMAIL_CONFIRM=true`: only once `[mailer]` works.
- `REGISTER_MANUAL_CONFIRM = true`: emergency switch that holds every new account until an admin
activates it.
- `[service] EMAIL_DOMAIN_ALLOWLIST` / `EMAIL_DOMAIN_BLOCKLIST`: restrict sign-up email domains.
- **Edge rate limit:** a rate-limit rule at your proxy or CDN on `POST /user/sign_up` and
`POST /user/login` (for example, a few requests per minute per IP). Gitea 1.23 has no built-in
request rate limit.
Known gaps:
- The sign-up and password-reset forms reveal whether a username or email is taken (email
enumeration).
- Names reserved by the hosted instance's `/join` flow are not reserved in Gitea itself.
### Rollout order (existing instance with a captcha)
1. Add the edge rate limit and verify it: a burst of `POST /user/sign_up` from one IP gets 429.
2. Then set `ENABLE_CAPTCHA = false` (and the settings above) and restart Gitea.
3. Rollback: set `ENABLE_CAPTCHA = true` and restart.
## Contributing

View File

@ -49,10 +49,31 @@ SECRET_KEY = ${SECRET_KEY}
INTERNAL_TOKEN = ${INTERNAL_TOKEN}
MIN_PASSWORD_LENGTH = 10
PASSWORD_CHECK_PWN = false
; Outgoing requests (webhooks, OAuth2 clients) may reach public hosts only; loopback / LAN
; targets need an explicit [security] ALLOWED_HOST_LIST entry. Blocks SSRF into your network.
EGRESS_MODE = lax
; Outgoing requests (webhooks etc.) may reach public hosts only: no loopback, private or
; link-local addresses (SSRF guard). Gitea >= 28 reads this here; [webhook] ALLOWED_HOST_LIST is
; deprecated there. On Gitea 1.23 put the same line under [webhook] instead.
ALLOWED_HOST_LIST = external
[packages]
; Package registry off: no unbounded blob uploads.
ENABLED = false
[actions]
; Actions off: no runner jobs or artifact storage.
ENABLED = false
[attachment]
; Issue/PR attachments capped at 4 MB, 3 files per upload.
ENABLED = true
MAX_SIZE = 4
MAX_FILES = 3
[repository.upload]
; Web-UI file uploads capped at 4 MB, 3 files.
FILE_MAX_SIZE = 4
MAX_FILES = 3
[oauth2]
JWT_SECRET = ${JWT_SECRET}
@ -61,11 +82,15 @@ JWT_SECRET = ${JWT_SECRET}
DISABLE_REGISTRATION = ${DISABLE_REGISTRATION}
SHOW_REGISTRATION_BUTTON = true
; Agent-native: no captcha on sign-up or login. Abuse is limited by the per-user
; defaults below (repo cap, no orgs, no migrations/mirrors, egress lax), not by a challenge.
; defaults (repo cap, no orgs, no migrations/mirrors, webhooks to external hosts only,
; no packages/Actions, small attachments, no password basic auth), not by a challenge.
ENABLE_CAPTCHA = false
REQUIRE_CAPTCHA_FOR_LOGIN = false
; API and git-over-HTTP accept tokens only; password basic auth is refused.
ENABLE_BASIC_AUTHENTICATION = false
; Only turn this on once [mailer] really delivers mail.
REGISTER_EMAIL_CONFIRM = ${REGISTER_EMAIL_CONFIRM}
; Emergency switch: set true to hold new sign-ups until an admin activates them.
REGISTER_MANUAL_CONFIRM = false
ALLOW_ONLY_EXTERNAL_REGISTRATION = false
; Public repos stay browsable without an account.

View File

@ -40,7 +40,8 @@ H=(-H "Authorization: token $T" -H "Content-Type: application/json")
curl -fsS "${H[@]}" -X POST "$A/user/repos" -d '{"name":"demo","auto_init":true,"default_branch":"main"}' >/dev/null && ok "admin created repo kitadmin/demo" || bad "create repo"
./kit.sh protect kitadmin/demo >/dev/null && ok "branch protection applied to main" || bad "protect"
git -c credential.helper= clone -q "http://kitadmin:$P@127.0.0.1:$PORT/kitadmin/demo.git" "$CLONE"
# Password basic auth is off (ENABLE_BASIC_AUTHENTICATION=false); git over HTTP uses the token.
git -c credential.helper= clone -q "http://kitadmin:$T@127.0.0.1:$PORT/kitadmin/demo.git" "$CLONE"
cd "$CLONE"; git config user.name e2e; git config user.email e2e@example.com
echo direct > direct.txt; git add direct.txt; git commit -qm direct
if git -c credential.helper= push -q origin main 2>/dev/null; then bad "direct push to main was ACCEPTED"; else ok "direct push to main rejected (even for the admin)"; fi
@ -66,6 +67,31 @@ made=1; for i in $(seq 2 11); do
[ "$c" = 201 ] && made=$((made+1))
done
[ "$made" = 10 ] && ok "repo limit: regular user stops at 10 repos" || bad "repo limit: user made $made repos"
c=$(curl -s -o /dev/null -w '%{http_code}' -H "Authorization: token $AT" -H "Content-Type: application/json" -X POST "$A/user/repos" -d '{"name":"r-eleventh"}')
[ "$c" = 403 ] && ok "11th repo refused (403)" || bad "11th repo: $c"
# --- adversarial checks (agent-native sign-up, no captcha) ---
c=$(curl -s -o /dev/null -w '%{http_code}' -u "kitadmin:$P" "$A/user"); [ "$c" = 401 ] && ok "password basic auth refused on the API (401)" || bad "password basic auth: $c"
c=$(curl -s -o /dev/null -w '%{http_code}' -u "kitadmin:$T" "$A/user"); [ "$c" = 200 ] && ok "token as basic-auth password still works" || bad "token basic auth: $c"
c=$(curl -s -o /dev/null -w '%{http_code}' -H "Authorization: token $T" "$A/user"); [ "$c" = 200 ] && ok "token header works" || bad "token header: $c"
(cd "$CLONE" && git checkout -qb pw-push && if git -c credential.helper= push -q "http://kitadmin:$P@127.0.0.1:$PORT/kitadmin/demo.git" pw-push 2>/dev/null; then exit 1; fi) && ok "git push over HTTP refuses password basic auth" || bad "git HTTP push accepted a password"
c=$(curl -s -o /dev/null -w '%{http_code}' -H "Authorization: token $AT" -H "Content-Type: application/json" -X POST "$A/orgs" -d '{"username":"e2e-org2"}'); [ "$c" = 403 ] && ok "agent token cannot create an org (403)" || bad "agent org create: $c"
c=$(curl -s -o /dev/null -w '%{http_code}' -H "Authorization: token $AT" -X PUT --data-binary 'blob' "http://127.0.0.1:$PORT/api/packages/e2e-agent/generic/p/1.0/f.bin"); [ "$c" != 201 ] && ok "package upload refused ($c, registry off)" || bad "package upload accepted"
LPORT=$((PORT+7)); HIT="$TMP/webhook-hit"
python3 -c "import http.server,sys
class H(http.server.BaseHTTPRequestHandler):
def do_POST(s): open(sys.argv[2],'w').write('hit'); s.send_response(200); s.end_headers()
def log_message(s,*a): pass
http.server.HTTPServer(('127.0.0.1',int(sys.argv[1])),H).serve_forever()" "$LPORT" "$HIT" & LPID=$!
sleep 0.5
HID=$(curl -s -H "Authorization: token $AT" -H "Content-Type: application/json" -X POST "$A/repos/e2e-agent/agent-repo/hooks" \
-d "{\"type\":\"gitea\",\"active\":true,\"config\":{\"url\":\"http://127.0.0.1:$LPORT/\",\"content_type\":\"json\"}}" | python3 -c 'import json,sys;print(json.load(sys.stdin).get("id",""))' 2>/dev/null || true)
[ -n "$HID" ] && ok "webhook to 127.0.0.1 created (id $HID); delivering test" || bad "webhook create failed"
[ -n "$HID" ] && curl -s -o /dev/null -H "Authorization: token $AT" -X POST "$A/repos/e2e-agent/agent-repo/hooks/$HID/tests"
sleep 4; kill $LPID 2>/dev/null || true
[ ! -e "$HIT" ] && ok "webhook to 127.0.0.1 not delivered (ALLOWED_HOST_LIST=external)" || bad "webhook reached 127.0.0.1"
grep -q 'not allowed\|blocked\|ALLOWED_HOST' "$TMP/home/log/"*.log 2>/dev/null && ok "Gitea logged the blocked webhook host" || echo "note: no block line found in the Gitea log (delivery absence is the check)"
echo "PENDING sign-up burst: throttling is the edge rate-limit rule (Cloudflare/proxy), not Gitea; test once the rule exists"
if grep -q "Deprecation" "$TMP/home/log/gitea.out"; then bad "config uses deprecated keys: $(grep -o 'config option `[^`]*`' "$TMP/home/log/gitea.out" | sort -u | tr '\n' ' ')"; else ok "no deprecated config keys in the Gitea log"; fi
echo; echo "$pass passed, $fail failed"