Search: nothing.this

The Reachy Mini daemon exposes an HTTP API for managing the robot. Its app installation endpoint, POST /apps/install in src/reachy_mini/daemon/app/routers/apps.py, has no authentication. The handler's only dependency is Depends(get_app_manager), which just hands back the manager object from application state, so nothing in the chain ever checks a credential. The endpoint takes an AppInfo body naming a Hugging Face Space. The daemon downloads that Space and installs it as a Python package through install_package in src/reachy_mini/apps/sources/local_common_venv.py, using uv or pip. Installing a Python package runs the package's own build and setup code, so whoever chooses the Space chooses what code the robot runs. Anyone can publish a public Hugging Face Space, so this is not a meaningful restriction on the attacker. How far this reaches depends on the model. In _resolve_bind_host in src/reachy_mini/daemon/app/main.py the daemon binds 0.0.0.0 when it runs as the wireless version and 127.0.0.1 otherwise, with the vendor's own comment explaining that the robot has to be reachable on the LAN. On a wireless unit, then, any host on the same network can install and run code on the robot without credentials. One related change has already shipped but does not fix this. Version 1.8.2 replaced the wildcard CORS policy with an allow list of localhost and Tauri origins. That closes the browser drive-by route, where a web page the victim visits silently calls the endpoint in the background. It has no effect on this issue: CORS is enforced by browsers and governs whether script may read a response, while a direct HTTP request from another machine on the network involves no browser, no preflight and no CORS check at all.

Pake turns a website into a desktop application built on Tauri. Every application it generates inherits two settings from the upstream template, and together they hand native functionality to untrusted web content. The first is in src-tauri/capabilities/default.json, which grants IPC access with "remote": { "urls": ["https://*.*"] }. That wildcard tells Tauri to accept IPC from any HTTPS origin, not just the site the application was built to wrap. The second is "withGlobalTauri": true in src-tauri/tauri.conf.json, which puts window.__TAURI__.core.invoke() in reach of ordinary page JavaScript. Tauri's access control list only checks plugin commands, the ones prefixed with plugin:. Commands the application registers itself through generate_handler!, known as app commands, are never checked against the ACL. So once an origin holds IPC access, it can call every app command with nothing else standing in the way. Pake registers download_file as an app command, and it does not appear in the permissions list because it does not need to. The practical effect is that any script running on any HTTPS page inside a Pake application can invoke the application's native commands. That includes third-party script the wrapped site loads on its own, such as analytics, advertising, or a compromised CDN. Anyone distributing their own Pake application gets the same access without asking for it. Chained with the path traversal in download_file that is tracked separately as CVE-2026-82635, this reaches arbitrary file write and persistent code execution.

Authentication Bypass by Capture-replay in ZenHive mpp allows an attacker holding a captured subscription activation credential to charge the payer repeatedly. The payer signs a Tempo KeyAuthorization over the chain id, key type, key id, expiry, limits and scopes only, with nothing tying it to the challenge that prompted it. MPP.Methods.Tempo.KeyAuthorization.verify/3 in lib/mpp/methods/tempo/key_authorization.ex pins each of those signed fields against the subscription request, and the access key it pins is a static per-endpoint server key, so one signed authorization verifies against every challenge the server issues for the same subscription terms. MPP.Methods.Tempo.Subscription.activate/4 deduplicates activations by challenge id, so presenting the captured credential under a fresh challenge produces a different dedup key, claim_activation succeeds, and the subscription transaction is built and broadcast again. Each replay charges the payer's wallet a new first-period settlement and re-authorizes the server key, bounded only by the subscription expiry and the chain's own semantics for re-installing an existing key. This issue affects mpp: from 0.14.0 before 0.16.2.

Apache Airflow: the Core API logout endpoint revokes only a session token presented as the _token cookie. When a client logs out presenting its credential as an Authorization bearer header instead, the endpoint returns its normal logout response but revokes nothing, so the token remains valid until it expires. An attacker who already holds a copy of that token keeps the victim's access after the victim has logged out and believes the session ended; the default token lifetime is 24 hours and is configurable. Affects API clients that authenticate with a bearer token rather than the browser session cookie. The attacker must already possess a copy of a valid token; obtaining one is outside the scope of this issue, and no privileges beyond the victim's own are gained. Users of apache-airflow are recommended to upgrade to apache-airflow version 3.3.2 or later, which fixes the issue.

### Summary `Repo.__init__` decides which directory is the git directory by testing candidate paths in an order that considers the real `.git` **last**. Two earlier tests can be satisfied by ordinary tracked files. Git reserves only the literal name `.git`, so `HEAD`, `objects/`, `refs/`, `config`, `gitdir`, `commondir` and `hooks/` at a repository root are all legal tracked content. Consequently, after a victim opens or clones an attacker's repository, GitPython resolves `git_dir` to the **working-tree root** while real git correctly resolves `<root>/.git`. Everything GitPython then treats as "inside the git directory" is attacker-authored content — including `hooks/`, which it executes. ### CVE-2026-87817 ### Affected code (3.1.59) The discovery loop in `git/repo/base.py` tests, in order: 1. `git/repo/base.py:299` — `isfile(curpath/gitdir)` **and** `isfile(curpath/commondir)` **and** `isfile(curpath/HEAD)` 2. `git/repo/base.py:320` — `is_git_dir(curpath)` 3. `git/repo/base.py:341` — `dotgit = osp.join(curpath, ".git")` ← the real git dir, considered last `is_git_dir` (`git/repo/fun.py:60`) requires only that `objects/` and `refs/` are directories and that `HEAD` is a file; **`HEAD`'s contents are never parsed.** The hook path is resolved from `index.repo.git_dir` (`git/index/fun.py:73`), i.e. the mis-resolved directory. ### Proof of concept Requires only `pip install GitPython==3.1.59`. Full script attached as `poc1_rce.py`; it runs entirely in a temp directory and the payload only writes a marker file. Attacker repository — four ordinary tracked files at the root: | path | mode | content | |---|---|---| | `gitdir` | 100644 | `.git\n` | | `commondir` | 100644 | `.git\n` | | `HEAD` | 100644 | `ref: refs/heads/master\n` | | `hooks/pre-commit` | **100755** | `#!/bin/sh` + payload | Victim — two ordinary calls: ```python repo = git.Repo.clone_from(url, dst) # or git.Repo(dst) repo.index.commit("automated commit") # code execution happens here ``` Observed on the PyPI release 3.1.59 (Linux and Windows): ``` real git says the git dir is : /tmp/.../victim/.git GitPython says it is : /tmp/.../victim <- shadowed attacker's hook executed : True git fsck : (clean) ``` The `pre-commit` hook runs at `git/index/base.py:1201`, before `write_tree()`, so it fires even though the commit later fails. ### Impact A service that opens or clones an untrusted repository with GitPython — a CI runner building a fork pull request, a code-scanning/SBOM service, a mirror, a dependency bot, an AI code-review/agent tool — can be made to: 1. **Execute arbitrary commands** via the tracked `<root>/hooks/pre-commit` when the victim calls `index.commit()`. 2. **Read files outside the repository**: the tracked `<root>/config` becomes the repository config and is parsed with `merge_includes=True` (`git/repo/base.py:765`), so `[include] path = ~/.aws/credentials` discloses the file. (`Repo._config_reader` still defaults `merge_includes=True`, so the hardening added in 3.1.59 for `.gitmodules`/GHSA-7833 does not cover this path.) 3. **Write a config file to an attacker-chosen directory** via an absolute tracked `commondir`. Delivery is silent: `git clone` exits 0, `git fsck` (including `--strict`, and with `transfer.fsckObjects`/`fetch.fsckObjects=true`) reports nothing, and the clone passes every read-only probe (`head.commit`, `branches`, `is_dirty()`, `untracked_files`, `iter_commits`) because a tracked `commondir` of `.git` pins `common_dir` to the real `.git`. ### Threat model / preconditions - Attacker controls the content of a repository the victim opens or clones with GitPython (public repo, fork PR, mirrored dependency). - For code execution, the victim performs a commit via GitPython's native `index.commit()`. For the file-read impact, opening the repo and reading config is enough. - `GIT_WORK_TREE` is **not** a mitigation — `git_dir` is assigned and the discovery loop breaks before the environment is consulted. - Real git is unaffected; only GitPython mis-resolves the directory. ### Remediation 1. Test `curpath/.git` **before** the `gitdir`/`commondir`/`HEAD` triple and before `is_git_dir(curpath)`. 2. Resolve `hook_path` / `_commit_hook_path` / `_get_validated_reflog_path` through `repo.common_dir`. 3. Containment-check `commondir`/`gitdir` contents before joining (reuse `SymbolicReference._get_validated_path`). 4. Validate `HEAD` in `is_git_dir` (require `ref: refs/...` or a 40-hex sha). 5. Pass `merge_includes=False` in `Repo._config_reader` (`git/repo/base.py:765`). ### Note for the maintainers Commit `406b98e1` (2026-05-31, "respect core.hooksPath for commit hooks") resolved the hook path via `git rev-parse --git-path`, which is immune because git rediscovers the true `.git`. Commit `9bc287a2` (2026-07-20) reverted it to avoid an unconditional `rev-parse` dependency — reintroducing the exec step. Measured at each commit with the healthy-lo

pgAdmin 4's Webserver authentication source is intended to accept an identity asserted by the web server or reverse proxy in front of pgAdmin, delivered through the WSGI/CGI environment. WebserverAuthentication.get_user() read config.WEBSERVER_REMOTE_USER from request.environ and, when that returned nothing, fell back to reading the same name directly from the inbound HTTP request headers via request.headers.get(). An inbound HTTP header is written by whoever sends the request, so any client able to reach pgAdmin could supply that header itself and be authenticated as any username it named, including an existing Administrator, without presenting a password or any other credential. The environment lookup could also be satisfied by a client-supplied header whenever WEBSERVER_REMOTE_USER was configured to an HTTP_-prefixed or hyphenated name such as HTTP_X_FORWARDED_USER or X-Forwarded-User, since WSGI servers place inbound headers into the environment under exactly those names. Deployments are affected only when 'webserver' is enabled in AUTHENTICATION_SOURCES. The fix distinguishes a genuine CGI/WSGI variable from a header-derived one and implicitly trusts only the former. A header-asserted identity is now accepted only when the operator explicitly opts in via WEBSERVER_REMOTE_USER_FROM_HEADER, the request arrives from a peer listed in WEBSERVER_TRUSTED_PROXIES, and, when configured, a shared secret supplied in WEBSERVER_SHARED_SECRET_HEADER matches WEBSERVER_SHARED_SECRET under a constant-time comparison. The trusted-peer check deliberately reads the real socket peer address rather than request.remote_addr, because ProxyFix rewrites the latter from the client-controlled X-Forwarded-For header and would otherwise allow an attacker to claim to be the trusted proxy. As defence in depth, login() now refuses any account whose auth_source is not 'webserver', so a misconfigured trust gate cannot be used to assume an internal or LDAP account. This issue affects pgAdmin 4: from 6.2 before 9.18.

Authentication Bypass by Spoofing vulnerability in team-alembic ash_authentication allows an attacker who operates one identity-provider connection of a dynamic_oidc strategy to be signed in as a local user established through a different connection. The strategy is meant to keep each connection in its own identity namespace by writing every UserIdentity row's strategy field as "<name>/<connection_id>", but that namespacing never takes effect. __connection_id__ is populated only on the ephemeral runtime struct built per request in dynamic_oidc/plug.ex, and DynamicOidc.IdentityChange.change/3 re-fetches the strategy from the compile-time DSL through Info.strategy_for_action, yielding the persisted struct whose __connection_id__ is its defstruct default of nil. OAuth2.identity_strategy_name/1 therefore falls back to the bare strategy name for both the identity write and the reads in oauth2/user_resolver.ex and oauth2/sign_in_preparation.ex. Since the identity resource's unique key is (uid, strategy), one row exists per sub across every connection, and the identity-match branch runs before any email check. Neither strategy handles iss, so nothing else distinguishes the issuers: OpenID Connect Core section 5.7 makes sub unique only within an issuer, so two connections numbering subjects independently share one subject space. This issue affects ash_authentication: from 5.0.0-rc.10 before 5.0.0-rc.14.

Improper Authentication vulnerability in team-alembic AshAuthentication allows an attacker holding a sign-in token for one authenticated resource to be signed in as a user of a different resource. AshAuthentication.Strategy.Password.SignInWithTokenPreparation.extract_primary_keys_from_subject/2 parses the JWT sub claim (for example user?id=1) with URI.parse/1 and keeps only its query string, discarding the path segment that names the subject the token was issued for. Nothing else restores that binding: AshAuthentication.Jwt.verify/3 checks the signature, exp, nbf, jti and the library-version claims, the purpose check only requires sign_in, and the remaining comparison is over primary-key field names, which are identical across resources. The WebAuthn sign-in and remember-me preparations carry copies of the same helper and drop the path in the same way. The magic link sign-in path pins the subject name against the resource and is not affected. This issue affects ash_authentication: from 3.10.5 before 4.15.0 and from 5.0.0-rc.0 before 5.0.0-rc.14.

Time-of-check Time-of-use (TOCTOU) Race Condition vulnerability in team-alembic AshAuthentication allows an attacker holding a leaked magic link to replay its single-use token and authenticate as the target subject. A magic link configured with single_use_token?, which is the default, is meant to be redeemable exactly once, but nothing serialises the token's validity check against its consumption, so concurrent redemptions of one token all succeed and each yields a full user token. Sign-in verifies the JWT with Jwt.verify/4 and revokes it only afterwards: AshAuthentication.Strategy.MagicLink.SignInPreparation revokes in a Query.after_action callback, and AshAuthentication.Strategy.MagicLink.SignInChange in an after_transaction hook that runs once the sign-in has already committed. AshAuthentication.TokenResource.Actions.revoke/3 writes the revocation as an upsert, so a concurrent duplicate revocation silently succeeds instead of conflicting and no request ever loses the race. This issue affects ash_authentication: from 3.9.0 before 4.15.0 and from 5.0.0-rc.0 before 5.0.0-rc.14.

Use of a One-Way Hash with a Predictable Salt vulnerability in team-alembic AshAuthentication allows readers of the audit store to recover the client IP addresses that the audit log add-on's :hash privacy mode is meant to pseudonymise. AshAuthentication.AddOn.AuditLog.IpPrivacy.hash_ip/1 computes a single unkeyed :crypto.hash(:sha256, salt <> ip) and truncates the result to 16 hexadecimal characters. The salt is read from the :audit_log_ip_salt or :secret application config keys, and falls back to the constant "default-salt-change-in-production" published in the library source when neither is set, with nothing warning that the default is in use. The IPv4 space is only 2^32 values and SHA-256 is fast, so the whole hash table is precomputable and every stored value maps back to its source address. Truncating to 16 characters does not help, and even a configured salt leaves the hash cheap enough to enumerate once it leaks. This issue affects ash_authentication: from 4.12.0 before 4.15.0 and from 5.0.0-rc.0 before 5.0.0-rc.14.

djust provides Phoenix LiveView-style reactive server-side rendering for Django with Rust-powered performance. Prior to version 1.0.7, `djust.mixins.model_binding.ModelBindingMixin` provides a default `update_model` event handler and is part of the LiveView base MRO, so every LiveView exposes it. It `setattr`s a view attribute whose name is client-supplied (`field`), gated only by: reject `_`-prefixed names; reject a 14-entry denylist of framework internals (`FORBIDDEN_MODEL_FIELDS`); optional `allowed_model_fields` which defaults to None = allow all; and `hasattr` existence. As a result, a client can set any public, existing view attribute — not just the fields actually bound with `dj-model=` in the rendered template. The denylist covers framework plumbing but nothing about developer business/authz state, and the allowlist is opt-in (off by default). A developer who binds one `dj-model="search"` input and also keeps `self.account_id` / `self.is_admin` / `self.total_price` as view state does not realize a client can set ALL of them via `{type:event, event:"update_model", params:{field, value}}` over the WebSocket. Type coercion matches the target attribute's type (so `"true"` -> bool True), aiding the attacker. This issue is fixed in djust 1.0.7. As a workaround, set `allowed_model_fields` explicitly on every view using dj-model (or subclassing LiveView) to the minimal list of bindable fields; do not keep authorization/ownership state in public view attributes that share the view with dj-model bindings.

Server-Side Request Forgery (SSRF) vulnerability in Slab safeurl allows an attacker who controls a validated URL to reach internal network destinations the library is configured to block. Only IPv4 addresses are matched against the reserved ranges and the blocklist. Every other address is treated as matching nothing, so a destination that is rejected in its IPv4 form is accepted when written as an IPv6 address, IPv6 entries in the blocklist never match, and a host that resolves to no IPv4 address is accepted regardless of where it points. Deployments that rely on the allowlist instead are unaffected, because there an unmatched address is rejected. This issue affects safeurl: from 0.1.0 onward.

## Summary API token permission checks only match the HTTP method and route path. The `expand` query parameter on task read endpoints embeds data from other permission groups (task comments, reactions, time entry counts) without checking whether the token holds those scopes. A token scoped only to `tasks` read permissions can therefore read task comments and reactions it was explicitly not granted. ## Details `models.CanDoAPIRoute` (`pkg/models/api_routes.go`, ~line 440) authorises API tokens purely by comparing `method` and `c.Path()` against the routes stored for each granted permission group. The query string is never consulted. The task read endpoints accept `expand`: - `GET /api/v1/tasks/:task`, `GET /api/v1/tasks` (`pkg/models/tasks.go` ReadOne/ReadAll, `pkg/models/task_collection.go`) - `GET /api/v2/tasks/:id`, `GET /api/v2/tasks`, `GET /api/v2/projects/:project/tasks` (`pkg/routes/api/v2/tasks.go`, `task_collection.go`) Accepted values include `comments`, `comment_count`, `reactions`, `time_entries_count`. `addMoreInfoToTasks` (`pkg/models/tasks.go`, ~line 683) loads and embeds that data. At that layer only a `web.Auth` (the plain owner user, resolved by `auth.GetAuthFromClaims`) is available, so the token's scopes cannot be enforced there either. Result: the `tasks_comments`, `reactions`, and `time_entries` permission groups are advisory for any data reachable through a task expansion. ## Impact A holder of an API token scoped to `tasks: [read_one]` or `tasks: [read_all]` (or `projects_views_tasks: [read_all]`) can read the full bodies of task comments, all reactions, and time entry counts on every task the token owner can access, despite `GET /api/v1|v2/tasks/:task/comments` and the reactions endpoints correctly returning 401 for the same token. The leak is limited to data the token owner can already see, and is read-only. The realistic victim is a user who grants a narrowly scoped token to a third-party integration expecting comments to stay private. ## Proof of Concept Against the test fixtures (`pkg/db/fixtures/api_tokens.yml`, token 1 has `{"tasks":["read_all","update"]}`, plaintext `tk_2eef46f40ebab3304919ab2e7e39993f75f29d2e`): ``` GET /api/v2/tasks/1/comments Authorization: Bearer tk_2eef46f40ebab3304919ab2e7e39993f75f29d2e -> 401 {"code":11,"message":"missing, malformed, expired or otherwise invalid token provided"} GET /api/v2/tasks?expand=comments&filter=id%3D1 Authorization: Bearer tk_2eef46f40ebab3304919ab2e7e39993f75f29d2e -> 200, response items[0].comments contains the full comment objects ``` Same behaviour with `expand=reactions`, and on the v1 endpoints `GET /api/v1/tasks?expand=comments` / `GET /api/v1/tasks/1?expand=comments`. Any user-created token with only `tasks` read permissions reproduces this. ## Recommended Fix Enforce expansions in the single existing choke point, `models.CanDoAPIRoute`: after the method/path match succeeds, read `c.QueryParams()["expand"]` and require the token to hold the owning group's read permission for each value, e.g. - `comments`, `comment_count` -> `tasks_comments.read_all` - `reactions` -> `reactions.read_all` - `time_entries_count` -> `time_entries.read_all` (`subtasks`, `buckets`, `is_unread` are task-level data and need nothing extra.) Doing this in the middleware covers both v1 and v2 without handler or model changes. Add a table-driven test alongside `pkg/webtests/api_token_method_matching_test.go` asserting a `tasks`-only token gets 401 with `expand=comments` and 200 once `tasks_comments.read_all` is added.

### Summary Once a user has TOTP enabled, the API still hands back the raw shared secret to anyone holding that account's access token. Reading it doesn't ask for the password, even though disabling TOTP does. So a stolen token, an XSS, or a browser left open is enough to copy the second factor into your own authenticator and keep generating valid codes indefinitely. ### Details `GET /api/v1/user/settings/totp` returns the full TOTP object, including the secret field and the otpauth:// provisioning URL. `GET /api/v1/user/settings/totp/qrcode` renders the same secret as a QR image. Neither requires re-authentication, the access token alone is enough. This is inconsistent with the rest of the flow: `POST /api/v1/user/settings/totp/disable` calls CheckUserPassword before it will turn TOTP off. So the destructive action is gated behind the password, but reading out the secret that backs the second factor isn't. There's also no reason for the secret to be readable at all once enrollment is finished, the client only needs it during setup. The fix is to stop returning secret/url/the QR code once enabled is true (only expose them during the enrollment window). Requiring the password on the read, like disable already does, would also be reasonable. ### PoC The attacker needs the victim's access token, from an XSS, a leaked/stolen token, or an unlocked session. TOTP must already be enabled on the account. 1. With the victim's bearer token, call: ``` GET /api/v1/user/settings/totp Authorization: Bearer <victim_token> ``` The response contains the shared secret and the otpauth:// URL: ``` {"secret": "<base32 secret>", "enabled": true, "url": "otpauth://totp/..."} ``` 2. Paste that secret (or scan `/api/v1/user/settings/totp/qrcode`) into any authenticator app. 3. It now produces the same 6-digit codes as the victim's device. Nothing is logged and the victim gets no notification. ### Impact This defeats the purpose of the second factor. 2FA is supposed to survive exactly this situation, a compromised password or a hijacked session, but a single read of the account's own settings hands over the shared secret with no step-up check and no trace. Combined with a known or later-recovered password, it's a durable account takeover that TOTP was meant to prevent.

### Summary When you share a project with a team, the API lets you attach any team on the instance, including teams you have nothing to do with, as long as you're an admin of the project. Listing a project's teams then returns each team's full member roster. So any logged-in user can spin up a throwaway project, attach team IDs one by one, and read back the name, description, and complete member list of every team on the instance. Normally you can only see teams you belong to; this path ignores that. ### Details Sharing a project with a team is a two-step flow: you PUT a team to the project, then you GET the project's teams to see them. The share endpoint only enforces that you're an admin of the target project, it doesn't care whether the team you're referencing is one you're allowed to see. Any existing team ID is accepted. The listing endpoint is even more permissive: read access to the project is enough, which you obviously have on a project you created. And its response isn't just a list of team names, it includes each team's description, creator, and a full members array with every member's username, display name, and admin flag. Put together, an attacker who owns a single project can walk the team ID space and pull the roster of every team in the instance, including teams they aren't a member of. This contradicts how team visibility works everywhere else in the product, where you only ever see teams you belong to. Emails are stripped from the roster, so what leaks is identities and org structure rather than contact details. The fix is to require that the caller actually has access to the team they're attaching, not just to the project. ### PoC The attacker needs nothing but an ordinary account. 1. Create a throwaway project you own: 2. Attach team IDs one by one (sweep the range you want): ``` PUT /api/v1/projects/<your_project_id>/teams {<SNIP>, "team_id": <N>, <SNIP>} ``` 3. Read them back with their rosters: ``` GET /api/v1/projects/<your_project_id>/teams ``` Each entry contains the team name, description, creator, and a members array with every member's username, display name, and admin flag, for teams you are not a member of. ### Impact Information disclosure of the instance-wide team directory: team names, descriptions, and full membership, for teams the attacker has no relationship with. This maps out the organisation's group structure and who belongs where, useful for targeting and social engineering. Any authenticated account is enough; no share link, no admin rights, no victim interaction.

### Summary Vikunja lets you export your data as a zip and import it back. The import doesn't limit how large the archive expands to, how many files it contains, or how much storage one user can consume, and it appears to hold the whole archive in memory while processing it. Because the archive is compressed, a ~20 MB upload made of highly compressible content expands to tens of gigabytes once imported. A single crafted upload is enough to exhaust the server's memory or fill its disk and knock the whole instance offline. It can be repeated or run in parallel, since nothing stops a user from starting several imports at once. ### Details The import accepts an archive in the same layout the built-in export produces: a manifest describing projects and tasks, plus the attachment files those tasks reference. Each individual file inside the archive is size-checked, but there's no ceiling on the total uncompressed size or on the number of files, and there's no per-user storage quota anywhere. On top of that, the server seems to buffer every attachment in memory before writing it out, so the peak memory cost is the sum of all decompressed files at once, not one at a time. Compression is the key factor: a file full of a repeating byte (e.g. zeroes) shrinks by roughly a thousand to one. So an archive that fits under the upload size limit (around 20 MB) can describe on the order of a thousand ~20 MB attachments, which the server then expands to tens of gigabytes in RAM and writes to disk. When memory runs out the process is killed; when the disk fills, the instance (and anything else on that host) stops working. Nothing serialises imports, so a user can also fire several in parallel to reach the limit faster. The fix is to cap the total uncompressed size and the file count per archive, stream each file to disk instead of buffering them all in memory, enforce a per-user storage quota, and refuse a new import while one is already running for that user. ### PoC The attacker needs an ordinary account that can use the import feature. 1. Build a zip in the export format containing one project with for example ~200 tasks, each referencing an attachment, and ~200 attachment files, each 18 MB of zero bytes. Filled with zeroes, the whole thing compresses to roughly 4 MB (under the upload limit). ```python #!/usr/bin/env python3 import argparse, json, os, zipfile CHUNK = b"\0" * (1024 * 1024) def build(out, n, size, version): tasks = [{"title": f"t{i}", "attachments": [{"file": {"id": i, "name": "a"}}]} for i in range(1, n + 1)] data = json.dumps([{"title": "bomb", "tasks": tasks, "child_projects": []}]).encode() with zipfile.ZipFile(out, "w", zipfile.ZIP_DEFLATED, compresslevel=9) as z: z.writestr("VERSION", version) z.writestr("data.json", data) for i in range(1, n + 1): with z.open(f"files/{i}", "w") as f: for _ in range(size): f.write(CHUNK) raw = n * size * 1024**2 print(f"[+] {out}: {os.path.getsize(out)/1024**2:.2f} MiB " f"({raw/1024**3:.1f} GiB uncompressed)") p = argparse.ArgumentParser() p.add_argument("-o", "--out", default="bomb.zip") p.add_argument("-n", "--files", type=int, default=200) p.add_argument("-s", "--size-mib", type=int, default=18) p.add_argument("--version", default="v2.5.0") a = p.parse_args() build(a.out, a.files, a.size_mib, a.version) ``` 3. Upload it to the import endpoint: ``` PUT /api/v1/migration/vikunja-file/migrate (multipart file upload: the crafted zip) ``` 4. Watch the server. Memory climbs into 3.5 GB within seconds and the disk fills with the expanded files. On a typical host the process is OOM-killed or the disk fills, and the instance stops responding for everyone. Repeat, or send several uploads at once, to reach the limit reliably. (You can derive the exact archive layout in step 1 by first exporting your own data once and inspecting the zip it hands back.) ### Impact Denial of service (availability). One authenticated user, with a single upload, can drive the server into tens of GB of memory and disk usage, killing the process or filling the disk. This takes the whole instance down for every user, and if the disk is shared with other services on the host, those go down too. No admin rights and no victim interaction are required, any account that can import data is enough. Recovery needs an operator to reclaim memory/disk and clear the partially written import.

### Summary The `/dav`, `/.well-known`, and `/feeds` groups are registered on the root Echo instance with only BasicAuth and no rate limiter. CalDAV BasicAuth accepts the plain account password, so password guessing over `/dav` is unbounded and never returns 429, while `/api/v1/login` is throttled from the tenth attempt. The only anti-brute-force control on the instance is therefore bypassable. ### Details `pkg/routes/routes.go` (~lines 238-249) registers `/.well-known`, `/dav`, and `/feeds` with `middleware.BasicAuth(...)` and nothing else; `registerCalDavRoutes` adds no limiter. `pkg/routes/caldav/auth.go` (~lines 88-93) falls through to `user.CheckUserCredentials` with the plain account password when no CalDAV token matches. In contrast, `/register`, `/login`, etc. are wrapped by `unauthRateLimit()` — an unconditional 10/min/IP pre-auth floor that ignores `ratelimit.enabled` (default false). TOTP-enabled accounts and bot users are correctly refused, so this targets password-only accounts. ### PoC (verified at runtime against v2.5.0) ``` POST /api/v1/login x25 wrong passwords -> 429 from attempt 2 (throttled) PROPFIND /dav/principals/{user}/ x60 wrong passwords -> 401 x60, 429 x0 GET /feeds/notifications.atom x30 wrong passwords -> 401 x30, 429 x0 PROPFIND /dav/... with correct password -> 207 (proves the 401s are real auth failures) ``` ### Impact The anti-brute-force floor guarding `/login` is entirely absent on `/dav`, `/feeds`, and `/.well-known`, giving an unbounded credential-guessing surface against account passwords. (bcrypt caps throughput to a few guesses/second, but nothing caps the number of attempts.) Reported as an authentication-control bypass, not a DoS. ### Fix Apply the unconditional pre-auth rate-limit floor to the `/dav`, `/.well-known`, and `/feeds` groups.

Description The setuid-root `worker-launcher` binary adjusts ownership and permissions of worker directories by walking the tree with FTS and calling `lchown` and `chmod` on each entry's full pathname while running with an effective uid of 0. Both syscalls re-resolve the path at the time of the call, after FTS has classified the entry, and the trees being walked are owned and writable by the untrusted topology user. A tenant running code on a supervisor node could therefore replace an intermediate directory component with a symbolic link between classification and the privileged operation, redirecting the root-owned `lchown` or `chmod` at an arbitrary file on the host. The operation is repeatable at will, since crashing a worker forces a relaunch and blob updates re-run the walk, so a failed attempt costs the attacker nothing. This crosses the boundary that `supervisor.run.worker.as.user` and container isolation are intended to enforce. It is the same defect class as the Hadoop container-executor issues from which this code derives. Mitigation Upgrade to 3.1.0, where the privileged walk operates on file descriptors it has already stat'd rather than on pathnames re-resolved at call time. Users who cannot upgrade immediately should not run untrusted topology code on supervisors configured with `supervisor.run.worker.as.user`, since the launcher is the boundary being crossed. Note that the launcher must be rebuilt and reinstalled after upgrading; replacing the Java artifacts alone is not sufficient. Credit The ASF -- found using Claude agents to study the security of open-source projects, validated and reported by Apache Storm.

Description `SimpleACLAuthorizer` evaluated the user-level command set by returning early when `nimbus.users` was empty, before `nimbus.groups` was considered. An operator who restricted cluster access by group alone, leaving `nimbus.users` unset, therefore received no restriction at all: every authenticated principal was permitted every user-level operation, including `submitTopology`, `beginFileUpload` and `getNimbusConf`. `docs/SECURITY.md` presents `nimbus.groups` as a supported way to lock down a cluster, so a deployment following the documentation could believe it was restricted while it was not. The failure is silent; nothing in the logs or the configuration indicates that the group list is being ignored. Both lists left empty continues to mean that no restriction is configured, which is the shipped default and is unchanged. Mitigation Upgrade to 3.1.0, where `nimbus.groups` is evaluated whether or not `nimbus.users` is set. Users who cannot upgrade immediately should additionally populate `nimbus.users` with the intended principals, since a non-empty user list causes the group list to be evaluated on affected versions. Operators should review Nimbus access logs for operations by principals outside the intended groups. Note that after upgrading, a cluster configured with `nimbus.groups` alone becomes restrictive for the first time. This includes `NimbusClient`, which calls `getLeader` on every connection, so clients outside the configured groups will begin to be refused. Credit The ASF -- found using Claude agents to study the security of open-source projects, validated and reported by Apache Storm.