### Summary Phoenix transports do not limit the number of channels in a given connection, making it easy to spawn hundreds of thousands of processes over a single connection, and, eventually reaching the max processes VM limit. The solution is to limit the number of channels per transport, so an attacker needs to start new HTTP/WebSocket connections, allowing third-party services to apply rate li

### Summary The Phoenix JavaScript presence client (`assets/js/phoenix/presence.js`) tests whether a presence already exists using a bare truthiness check (`state[key]`) rather than an own-property check. Because applications commonly track presences under a client-supplied username or id, the presence key can be attacker-controlled. A user who joins a channel and picks a key that names an `Objec

### Summary AshAuthentication's OAuth2 and OIDC family strategies matched the local user by email address rather than by the OpenID Connect `iss`/`sub` claim combination. A provider login presenting a victim's email (including an unverified, reused, or `email_verified: false` account) resolved to and signed in as the victim's existing local account. An unauthenticated attacker who can register an

### Summary An unauthenticated remote peer can crash any gRPC server built on this library by sending a small gzip-compressed frame that decompresses to gigabytes, exhausting the BEAM node's heap and triggering an OOM kill (denial of service). Introduced in https://github.com/elixir-grpc/grpc/commit/beae6800fc8baf126f3fe7107d86a50e105275ba ### Details `GRPC.Compressor.Gzip.decompress/1` (lib/grp

### Summary `'Elixir.GRPC.Server.Adapters.Cowboy.Handler':read_full_body/3` accumulates every received chunk into a single growing binary with no size cap. When the client omits the `grpc-timeout` header, the read timeout resolves to `:infinity`, allowing a slow-trickle attacker to hold the connection open indefinitely while memory grows. A single unauthenticated connection is sufficient to exhau

### Summary In the HTTP-to-gRPC transcoding layer of the `grpc` Hex package, query-string and request-body parameters can silently overwrite path-bound fields when building the decoded protobuf request struct. An authenticated attacker who can reach a transcoded endpoint can substitute any path-bound identifier (e.g. `user_id` from `/users/{user_id}/profile`) with an arbitrary value, bypassing au

### Summary `GRPC.Codec.Erlpack.decode/2` calls `:erlang.binary_to_term/1` directly on the raw gRPC message body without the `:safe` option. Any unauthenticated peer that can reach a gRPC endpoint with `Content-Type: application/grpc+erlpack` can crash the entire BEAM node via atom table exhaustion or, if a decoded fun term flows into a call site that invokes it, achieve remote code execution ins

### Summary `membrane_mp4_plugin` interns every 4-byte MP4 box name as a BEAM atom while parsing container headers, with no validation against an allow-list. Any code path that calls `Membrane.MP4.Container.parse/1` (or the bang variant) on attacker-controlled MP4 bytes can be made to exhaust the BEAM atom table, which is uncapped at the source but bounded by the runtime ceiling (around 1,048,576

### Summary Req's multipart form encoder interpolates the per-part `name`, `filename`, and `content_type` directly into the part headers without escaping. An attacker who can influence any of those values can inject CRLF-separated header lines, smuggle additional form fields, or prepend a whole extra part into the request the victim service sends downstream. ### Details `Req.Utils.encode_form_p

### Summary Req's default response pipeline auto-decodes archive and compressed bodies based on the server-supplied `content-type` (or URL extension) and materialises the full decompressed contents in memory with no size cap. An attacker who controls (or can redirect a victim into) an HTTP endpoint reached by `Req.get!/1` can return a tiny "decompression bomb" that expands to many gigabytes on th

### Summary Unbounded recursion depth in `Protobuf.Decoder` (Hex package `protobuf`, versions `>= 0.8.0, >` (tag `0x0A` = field 1, wire type 2). Use an iolist with a running byte-size to keep generation O(depth). 3. POST the body (a few MB at `depth = 1_000_000`) as `application/x-protobuf` to any endpoint that calls `Tree.decode/1`. 4. The non-tail `decode(bin, type)` in `value_for_field/3` re-e

### Summary `Tesla.Multipart.part_headers_for_disposition/1` interpolates `Content-Disposition` parameter values (field name, filename, and other opts) verbatim into the part header line without encoding or escaping any special characters. An attacker who controls a filename, field name, or other disposition parameter can use unescaped double-quotes to inject extra disposition key-value pairs, or

### Summary In the Mint adapter for the Tesla HTTP client library, `Tesla.Adapter.Mint.open_conn/2` passes the URL scheme of every outgoing request through `String.to_atom/1` with no allow-list validation. Because BEAM atoms are permanent (never garbage-collected) and the atom table is bounded at roughly 1,048,576 entries, an attacker who can vary the URL scheme across requests can mint one fresh

### Summary `Tesla.Middleware.FollowRedirects` is meant to strip the `Authorization` header when following a cross-origin redirect, but performs the check with a case-sensitive comparison against the lowercase string `"authorization"`. Because Tesla preserves header keys exactly as supplied by the caller, any application that sets the header with its RFC 7235 canonical casing (`"Authorization"`)

### Summary Any Tesla client pipeline that includes `Tesla.Middleware.DecompressResponse` or `Tesla.Middleware.Compression` eagerly decompresses HTTP response bodies with no size limit. A server under attacker control (or reached via a redirect) can return a tiny gzip-encoded payload that expands into gigabytes of BEAM heap, crashing or freezing the calling process. Stacking multiple `content-enc

### Summary `Tesla.Multipart.add_content_type_param/2` appends caller-supplied strings to the multipart `Content-Type` header with no validation. A param value containing `\r\n` splits the header line, allowing an attacker who controls any content-type parameter (charset, boundary parameter, etc.) to inject arbitrary headers into the outbound HTTP request. ### Details `add_content_type_param/2`

### Summary Mint's HTTP/2 client accepts `PUSH_PROMISE` frames from any server it connects to and inserts every promised stream into a per-connection map without consulting `max_concurrent_streams`. A malicious or compromised HTTP/2 server can flood the client with `PUSH_PROMISE` frames and withhold the matching response `HEADERS`, pinning one map entry per frame indefinitely until the client pro

### Summary Mint's HTTP/2 client accumulates `CONTINUATION` header-block fragments into a per-connection buffer with no cap on size or frame count. A malicious or compromised HTTP/2 server can drive the client's memory to arbitrary size by streaming an endless chain of `CONTINUATION` frames after a `HEADERS` frame that omits `END_HEADERS`, causing memory exhaustion and BEAM process death. A singl

### Summary Mint's HTTP/1 client accepts `Content-Length` header values with a leading `+` sign (e.g. `+0`, `+123`), which RFC 7230 forbids (`Content-Length = 1*DIGIT`). On a connection shared with a strict fronting proxy or load balancer, this parser disagreement is a response-smuggling primitive: the proxy frames the body one way, Mint frames it another, and bytes meant for one response leak in

### Summary Mint's HTTP/1 request encoder splices the caller-supplied `method` and `target` directly into the request line without character validation. An application that forwards attacker-controlled input as the HTTP method or the target to `Mint.HTTP.request/5` is exposed to request-line CRLF injection, allowing the attacker to terminate the request line early, inject arbitrary headers, and p

### Impact The QUIC client did not authenticate the server during the TLS 1.3 handshake. The CertificateVerify signature was not checked, the certificate chain was not validated, and the hostname was not compared against the certificate, so `verify` was effectively a no-op on the client. A man-in-the-middle on the network path could present any certificate and impersonate any server, defeating th

### Summary `oban_web` 2.12.0 through the current unpatched release exposes a `save-job` LiveView event handler that performs no authorization check, allowing any authenticated user (including those with `:read_only` access) to overwrite a queued job's `worker` field with any other `Oban.Worker` module present in the application. On the job's next execution attempt, Oban dispatches `perform/1` on

### Summary `oban_web` 2.12.0 introduced a cron expression parser that expands `-`-separated ranges without validating the endpoints. An attacker with access to schedule cron jobs can submit a malicious expression; when any user with dashboard access views the cron job list, `Oban.Web.CronExpr.describe/1` is called to render it, triggering allocation of gigabytes of memory and stalling or crashin

### Summary `ExAws.SNS.verify_message/1` fetches the signing certificate from the `SigningCertURL` field of the incoming SNS message without validating that the URL uses HTTPS or that its host is an AWS-owned SNS certificate domain. An unauthenticated attacker who can POST to any endpoint that calls `verify_message/1` can supply an attacker-controlled `SigningCertURL`, sign a forged SNS message w

### Summary [CVE-2026-47067](https://nvd.nist.gov/vuln/detail/CVE-2026-47067) is an atom table exhaustion vulnerability (CWE-770) in hackney's URL parser (`src/hackney_url.erl`). `hackney_url:parse_url/1` converts every URL scheme it encounters into a BEAM atom via `binary_to_atom/2`. Because BEAM atoms are never garbage-collected and the atom table has a hard limit of 1,048,576 entries, an attac

### Summary The WebSocket client in `src/hackney_ws.erl` imposes no upper bound on memory consumption across three distinct code paths. In each case, an attacker-controlled WebSocket server can exhaust the connecting process's memory without any authentication or special client configuration. ### Details **1. Handshake response buffer (`read_handshake_response/3`)** The function accumulates re

### Summary CRLF injection in hackney's WebSocket upgrade request builder (`src/hackney_ws.erl`). `init/1` copies the `host`, `path`, `headers`, and `protocols` options from the caller-supplied opts map verbatim into `#ws_data{}`, and `do_handshake/1` splices them directly into the raw HTTP/1.1 upgrade request by binary concatenation with no `\r\n` or `\0` stripping. A caller that passes any of t

### Summary `hackney_url:make_url/3` passes the URL query component directly into the HTTP/1.1 request target without percent-encoding `\r` or `\n`. RFC 3986 §3.4 requires characters outside the query grammar to be percent-encoded, but no validation or escaping occurs. An attacker who controls any portion of the URL passed to hackney can inject raw CRLF sequences into the request line, enabling H

### Summary `hackney_h3:await_response_loop/6` in `src/hackney_h3.erl` accumulates the HTTP/3 response body in memory without any size cap. The `after Timeout` clause is a per-message inactivity timer, not a wall-clock deadline: every received `stream_data` chunk, housekeeping `select` message, or `settings` frame resets it. A malicious HTTP/3 server that drips one small chunk every `Timeout - 1`

### Summary The HTTP/3 redirect handler in `src/hackney_h3.erl` forwards the original request headers (`Authorization`, `Cookie`, `Proxy-Authorization`) and, for 307/308 responses, the original request body to the redirect target without checking whether the target host matches the origin. When `follow_redirect` is enabled and a server responds with a cross-origin `Location`, hackney delivers the

### Summary `hackney_url:normalize/2` URL-decodes the host component of a parsed URL, but the caller's SSRF allowlist runs before normalization using OTP's `uri_string:parse/1` and `inet:parse_address/1`, neither of which decodes percent-escapes in hostnames. A URL like `http://%31%32%37%2E%30%2E%30%2E%31/` presents an encoded, non-IP-looking host to the validator, which passes the allowlist chec

### Summary CRLF injection in `hackney_cookie:setcookie/3` (`src/hackney_cookie.erl`). The function validates `Name` and `Value` against CR/LF and control characters but concatenates the `domain` and `path` options verbatim into the output binary. If either option carries attacker-controlled data, a `Host` header forwarded as the cookie domain, a request URI forwarded as the cookie path, a `\r\n`

### Summary The SOCKS5 transport in `src/hackney_socks5.erl` correctly applies the caller-supplied timeout to the SOCKS5 negotiation phase, but then upgrades the tunnel to TLS using `ssl:connect/2` (the two-argument form), which defaults to `infinity`. The `Timeout` value is in scope at that call site but is never forwarded. A hostile or man-in-the-middled SOCKS5 proxy that completes the SOCKS5 h

### Summary [CVE-2026-47066](https://nvd.nist.gov/vuln/detail/CVE-2026-47066) is an infinite loop (CWE-835) in hackney's Alt-Svc response header parser (`src/hackney_altsvc.erl`). When an HTTP server returns an `Alt-Svc` header whose value begins with a non-token byte (e.g. `!`, `@`, `=`, `;`), the parser enters a tight tail-recursive loop that pins an Erlang scheduler at 100% CPU and permanently

### Summary The storybook iframe LiveView accepts a PubSub topic from the URL query string and broadcasts its own pid onto that topic with no check that the topic belongs to the current session. Any unauthenticated visitor who knows or guesses another user's playground topic can hijack the playground↔iframe handshake, causing the victim's playground to send its control messages to an attacker-cont

### Summary An attacker who can deliver `psb-assign`, `psb-toggle`, `psb-set-theme`, `upper-tab-navigation`, `lower-tab-navigation`, `playground-change`, or `playground-toggle` LiveView events to a mounted Phoenix Storybook playground can flood the BEAM atom table with attacker-controlled strings, permanently leaking atoms until the VM hits its ~1,048,576 atom ceiling and crashes the entire node.

### Summary An unsafe HEEx template generation vulnerability allows any unauthenticated user to execute arbitrary code on the server. The phoenix_storybook playground accepts user-controlled attribute values over WebSocket and interpolates them unsanitized into a HEEx template that is subsequently compiled and evaluated with full Elixir `Kernel` access. ### Details The vulnerability is a three-st

### Summary An Allocation of Resources Without Limits or Throttling vulnerability in `Plug.Conn.read_part_headers/2` allows an unauthenticated attacker to exhaust server memory by sending a crafted `multipart/form-data` request, causing a denial of service. ### Details `Plug.Conn.read_part_headers/2` in `lib/plug/conn.ex` does not obey its `:length` parameter. There is no upper bound on the siz

### Summary SQL injection in `Postgrex.Notifications.listen/3`: the `channel` argument is interpolated straight into `LISTEN "..."` / `UNLISTEN "..."` without escaping the `"` character. Any caller that lets a user influence the channel name (e.g. a pub/sub bridge that uses a tenant id or topic slug as the channel) and the name is not sanitized can execute arbitrary SQL on the notifications conne

Summary `decimal` doesn't bound the exponent on parsed input, so something like `"1e10000000"` is parsed fine but then explodes the memory to more than 7GB if you run e.g. `Decimal.add(Decimal.parse("1e10000000"), 1)` because for positive `exp`, the function tail-recurses with `coef * 10` and `exp - 1` per iteration, growing the bignum coefficient by one digit each step. In the worst case, one req

### Summary An unauthenticated denial-of-service vulnerability in Phoenix's long-poll transport allows a remote client to allocate a large amount of memory with a HTTP request. A handful of concurrent requests can be sufficient to let the node run out of memory. See also https://cna.erlef.org/cves/CVE-2026-32689.html. ### Details The unoptimised code path exists on the `application/x-ndjson` P

### Summary Bandit's HTTP/2 parser checks frame size *after* it has already buffered the full body, instead of when it sees the 9-byte header. A peer can announce a 16 MiB frame on a connection that agreed to 16 KiB frames and the server will silently buffer up to 1024× the agreed budget per connection. Across many connections this becomes a memory-pressure DoS. Severity: medium. ### Details In

### Summary Bandit reflects the client-supplied URI scheme into `conn.scheme` without verifying the actual transport. Over a plaintext HTTP/1.1 connection (or h2c), an unauthenticated attacker can send an absolute-form request target like `GET https://victim/path HTTP/1.1` and the application observes `conn.scheme = :https` even though no TLS was negotiated. Any downstream Plug logic that trusts `

### Summary Bandit is vulnerable to CL.CL HTTP request smuggling: it silently accepts requests with two `Content-Length` headers whose values differ, takes the first value, and dispatches the body bytes as a second pipelined request on the same keep-alive connection. RFC 9110 §5.3 prohibits multiple lines for singleton fields like `Content-Length`, and RFC 9112 §6.3 item 5 requires the recipient

### Summary A single unauthenticated WebSocket client can exhaust server memory in any Bandit-fronted application that accepts WebSocket connections. The fragmented-message reassembly path appends every `Continuation{fin: false}` frame's payload to a per-connection iolist with no cumulative size cap, so a peer that streams continuation frames indefinitely (never setting `fin=1`) grows BEAM heap li

### Summary When a Bandit-fronted server has explicitly enabled WebSocket permessage-deflate (`compress: true`), an unauthenticated client can OOM the BEAM with a single ~6 MiB WebSocket frame. Bandit's inflate step has no output-size cap, so a small high-ratio compressed frame (e.g. zeros, ~1024:1 ratio) decompresses unbounded into the connection process before any application code runs. Phoenix

## Summary An unauthenticated remote denial-of-service vulnerability in `Plug.Cowboy.Conn` allows any attacker who can reach an HTTPS Plug.Cowboy listener via HTTP/2 to permanently exhaust the BEAM atom table and crash the entire Erlang VM. ## Am I Affected? All users running plug_cowboy with HTTP/2 may be affected, this includes Phoenix applications. If another HTTP adapter such as Bandit is u