.gitbay/wiki/Threat-Model.org

  1#+title: gitbay threat model
  2
  3What the forge trusts, what it refuses to do, and where the boundaries
  4are. This is the reference for security review; it complements the audit
  5log and hardening notes in [[Admin]].
  6
  7* What gitbay never does
  8
  9- *Execute repository content.* Git object contents are never run. Hooks
 10  are gitbay's own binary, invoked by git; they compute facts and ask the
 11  daemon over a unix socket. Repo files are only ever read.
 12- *Hold a signing key.* There is no server-side signing key. "Verified"
 13  means a signature made by a key the *user* registered — the server
 14  never vouches for a commit it did not receive already signed. Merge
 15  commits the server creates are honestly =unsigned=.
 16- *Serve repository HTML on its own origin as active content.* Raw file
 17  serving is =text/plain= with =nosniff=. Rendered markdown/org is
 18  sanitized (bluemonday) and served under a CSP that forbids scripts.
 19- *Put secrets in argv, URLs, or logs.* Import and mirror credentials,
 20  registration invites, and API tokens travel on stdin or in request
 21  bodies, never as command arguments (visible in =/proc=) or query
 22  strings. Tokens are stored only as SHA-256 hashes.
 23- *Name a mail recipient in the log.* A queued mail is logged by its
 24  queue row id, never by address, and the relay's own error is redacted
 25  before it is logged because a rejection usually quotes the address it
 26  rejected. The unredacted error and the address stay on the row, which
 27  an instance admin reads on =/admin=: the database holds who, the log
 28  holds which. One rule for every mail type — notifications, dependency
 29  reports and login links all drain the same queue (krz/gitbay#173).
 30- *Confirm the existence of private repositories.* Every surface answers
 31  "not found" identically for a private repo and a nonexistent one — web
 32  pages, git transport, control commands, release asset downloads.
 33
 34* Trust boundaries
 35
 36- *SSH public key = identity.* The SSH username is ignored; the presented
 37  key's fingerprint resolves to an account. Key uniqueness is global.
 38- *Per-instance trust.* Email verification and key registration are local
 39  to an instance and never transfer. Account migration re-registers keys
 40  and re-verifies emails on the target by design.
 41- *The control plane is one authenticated channel* (SSH), fully usable
 42  from stock OpenSSH. The JSON API fronts the same command registry with
 43  bearer tokens minted only over SSH; git transport never runs over it.
 44- *Anonymous surfaces* — HTTPS clone of public repos, =git://= where
 45  enabled, the read-only web UI — carry no credentials and expose only
 46  public data. HTTP push is refused via a pkt-line =ERR=, never a 401.
 47
 48* Attacker-controlled parsers
 49
 50Every parser that eats bytes from a pusher, a key registrant, or an
 51anonymous client has a fuzz target and must never panic:
 52
 53- =internal/protocol= — the SSH command tokenizer (fuzzed against argv
 54  round-tripping).
 55- =internal/gitd= — the =git://= pkt-line reader.
 56- =internal/sig= — the commit parser, the SSHSIG armor decoder and blob
 57  parser, and the OpenPGP armored-key reader.
 58
 59Run =deploy/audit.sh= to exercise them plus =go vet= and =govulncheck=.
 60
 61* Secret handling
 62
 63Tokens (web sessions, login links, email verification, API bearer tokens,
 64deploy/CI) are random 256-bit values. Only their SHA-256 hash is stored,
 65and verification is a database index lookup on that hash — the secret
 66itself is never compared in Go, so there is no timing oracle to exploit.
 67Webhook payloads are signed outbound with HMAC-SHA256; the forge verifies
 68no inbound HMAC.
 69
 70* Network-facing request forgery
 71
 72Anything that makes the *server* open an outbound connection to a
 73user-supplied address — webhook delivery, GitHub-history import
 74=--api-base=, mirror remotes — passes the same SSRF guard: the scheme
 75must be http/https and, unless =webhooks.allow_local= is set, the
 76resolved address must not be loopback, private, or link-local. The
 77webhook dialer re-checks at connect time so a DNS answer that changes
 78after validation still cannot reach private space. Redirects are never
 79followed.
 80
 81* Rendering pushed markup
 82
 83Rendered markup is attacker-controlled: a README, a wiki page and a
 84profile's about text are all whatever someone pushed or typed. The risk
 85is not only what the output contains but what the *parser* is willing to
 86go and fetch — the filesystem counterpart of the SSRF guard above.
 87
 88Org is rendered by go-org, whose default configuration resolves
 89=#+INCLUDE:= and =#+SETUPFILE:= targets with =os.ReadFile=. Both are
 90refused outright (=orgConfig()= in =internal/httpd=): the file is never
 91opened and the keyword stays the inert text it already was, so the rest
 92of the document renders normally. There is no safe subset to allow
 93instead — an absolute path skips go-org's relative-path join, a relative
 94one resolves against the daemon's working directory, and the content
 95came from a git object rather than a checkout, so there is no directory
 96to scope a read to. Markdown is goldmark, which has no include
 97mechanism. go-org's parse warnings are discarded rather than logged, so
 98pushed content cannot write to the server's log.
 99
100Org output is then sanitized (bluemonday UGC policy) because go-org
101passes raw HTML through — export blocks and inline export snippets —
102while goldmark drops it and needs no pass. The policy admits chroma's
103short token classes and nothing else.
104
105* Web responses
106
107Every response carries =Content-Security-Policy= (no scripts, no plugins,
108no embedding; inline styles allowed for chroma and label chips; images
109from any origin so external README images render), =X-Frame-Options:
110DENY=, =X-Content-Type-Options: nosniff=, =Referrer-Policy: no-referrer=,
111and =Strict-Transport-Security= when TLS is on. The UI needs no
112JavaScript, so =script-src 'none'= costs nothing.
113
114* The CI runner
115
116=gitbay-runner= is the one component that executes repository content.
117gitbayd never does: it reads =.gitbay/ci.yml= and queues a build, and a
118runner, polling over SSH, clones the commit and runs its steps.
119
120- *What the runner holds.* A key added with =keys add --scope runner=,
121  which the dispatcher confines to =runner next=, =runner log= and
122  =runner done= and to read-only git. A step that reads the key off the
123  disk gets exactly that: it cannot administer the instance, push, or
124  read a repository the runner's account cannot. An admin key still
125  works for the runner protocol so an operator can rotate at their own
126  pace; a runner host should not hold one.
127- *What a build sees.* The commit, the =GITBAY_*= variables and the
128  repository's secrets — unless the head came from another repository.
129  A merge request from a fork is built in the target as untrusted, with
130  no secrets, so a stranger's branch cannot read the target's deploy
131  credentials. The step environment is *constructed*, not inherited: a
132  build gets =PATH=, =HOME=, =LANG=, =CI=, its own variables and its
133  secrets, and nothing the operator set on the service. =HOME= is a build
134  home under the runner's =-workdir=, not the runner's own home, so a
135  build cannot read the =.netrc=, =.npmrc= or =.gitconfig= where tools
136  keep credentials. That home is shared by every build on the runner —
137  one build can poison a cache another reads, which is no more than
138  anything a step can already do as this user, and is what isolation
139  (krz/gitbay#144) is for.
140- *Where it runs.* Steps run in a rootless podman container, one per
141  job, with the workspace bind mounted and nothing else. The clone
142  happens outside it with the runner's key, so the container never sees
143  =GIT_SSH_COMMAND=, the key, or the runner's environment. =-isolation
144  none= runs steps on the host as before, for an instance where every
145  repository is trusted; there is no automatic fallback to it — a runner
146  configured for podman that cannot find one refuses to start, because
147  dropping isolation silently is worse than a stopped runner. The
148  systemd drop-in still adds =ProtectSystem=full= and the cgroup
149  protections, and =-repos= still limits a runner to named
150  repositories. =ProtectKernelTunables= is *off*: it overmounts =/proc=
151  in the unit's namespace and the kernel then refuses a proc mount in
152  any child user namespace, which every rootless container needs; the
153  container masks the same paths for the build itself.
154  =NoNewPrivileges= is *off*: rootless podman sets up its
155  namespace with the setuid =newuidmap=, which that flag blocks, so the
156  choice is between it and containers at all. Containers are the stronger
157  boundary — the flag constrained a process that was already running
158  arbitrary repository code, and under podman that code no longer runs in
159  the runner's process context. Under =-isolation none= there is no
160  container and the flag should be on.
161- *Images are provisioned by the operator, not fetched by a build.* The
162  runner passes =--pull=never=, so =image:= chooses among what the host
163  already has rather than naming anything on the internet. On an
164  instance with open registration that is the difference between a
165  curated set and arbitrary code from a registry nobody vetted.
166
167Under =-isolation none=, anything a step can do as the runner's user a
168pushed =ci.yml= can do. Under podman a step is confined to its
169container and the bind-mounted workspace, but the build home's caches
170are shared between builds, so one build can still leave something a
171later build reads. Treat the runner host as executing untrusted code:
172keep it off the daemon's host where the database lives, or scope it to
173repositories whose writers you trust.
174
175* What has not been audited
176
177The 2026-09 sweep (krz/gitbay#149) read this repository against the
178claims above. Its two findings — review verdicts from accounts without
179write access deciding merge gates (#147), and control commands having no
180write rate limit while the API had one (#148) — are fixed, as is #173,
181found afterwards in the same milestone. What it did not reach, and why,
182is recorded here rather than in a closed issue:
183
184- *The host.* systemd sandboxing, sshd on 2222, the firewall,
185  unattended-upgrades, fail2ban, disk and file modes, restic append-only
186  credentials. Reading production configuration is a separate exercise
187  from auditing the source, and needs doing against bay1.
188- *The supply chain beyond =govulncheck=.* No review of what the
189  dependency set is, who maintains it, or what a compromised release of
190  any of it would reach.
191- *The runner host as an execution environment.* Nobody has tried to
192  escape what is there. #144 covers the missing isolation.
193- *Timing and traffic analysis.* Token comparison is a hash index lookup
194  by design, but nothing has been measured.
195- *Denial of service by resource exhaustion* beyond rate: large pushes,
196  pathological diffs, deep histories, zip bombs in LFS.
197
198A sweep is a point in time. This section says what a reader should not
199assume has been checked.
200
201* Residual risks, accepted
202
203- External images in rendered READMEs and profile about text load from
204  their origin (no image proxy), which the author can use as a tracking
205  pixel against a viewer. A profile is the wider surface of the two: it
206  is linked from every commit and issue its owner touches. Documented;
207  proxying is future work.
208- Backups are consistent per the DB-snapshot-first ordering but are not a
209  single atomic snapshot; a few orphaned git objects are possible and
210  harmless (see [[Admin]]).
211- A global signature-verification epoch over-invalidates the cache on any
212  trust-input change. Correct, not a leak; a performance tradeoff.
213- A build's secrets are environment variables inside its container, so
214  they are visible to =podman inspect= as the runner's user — the same
215  user that already holds them in memory. They reach podman through a
216  0600 env file rather than argv, since =/proc= is world-readable.