CVE-2026-115913
vLLM is an inference and serving engine for large language models. Prior to 0.32.0, the Phi-4 multimodal audio encoder in phi4mm_audio.py an
CVSS
6.5
Medio
EPSS
—
KEV
—
Exploit Today
—
0-100
Publicado: 26 oct 2026 · Última mod.: 26 oct 2026 · CWE-770
Sin historial EPSS suficiente todavía.
vLLM is an inference and serving engine for large language models. Prior to 0.32.0, the Phi-4 multimodal audio encoder in phi4mm_audio.py and phi4mm_utils.py uses eager attention by default, allowing a remote client to submit one legal long audio clip that causes MultiHeadedAttention.forward_attention to materialize and retain full attention matrices in self.attn. The scaled dot-product attention branch does not materialize these matrices. The resulting quadratic GPU-memory use can cause an out-of-memory condition and terminate EngineCore for all users. This issue is fixed in version 0.32.0.
- github.comhttps://github.com/vllm-project/vllm/commit/3d52578e13454de0bb8352cac988347cdcc89928
- github.comhttps://github.com/vllm-project/vllm/pull/60614
- github.comhttps://github.com/vllm-project/vllm/releases/tag/v0.32.0
- github.comhttps://github.com/vllm-project/vllm/security/advisories/GHSA-8397-8qf6-f2gm
CVECVSSEPSSKEVRExplotTítuloVis.
CVE-2026-1159116.5 MED—
———vLLM is an inference and serving engine for large language models. From 0.19.0 until 0.32.0, authenticated callers can supply a request chat template when no server-side --chat-template is configured, or a malicious model repository can provide a template that is compiled and rendered by tokenizer.apply_chat_template() without a wall-clock timeout, output-size limit, or iteration bound. Rendering runs in a ThreadPoolExecutor with one worker by default, so a nonterminating render can block /v1/responses and other chat-template users after the request ends, while an output-expanding render can exhaust server memory. Harmony models are not affected because they do not use this Jinja path. This issue is fixed in version 0.32.0.4hCVE-2026-97200——
———Allocation of Resources Without Limits or Throttling vulnerability in hackney_conn in benoitc hackney allows a malicious HTTP server to exhaust client memory via an oversized or highly compressed response body. hackney:request/5 and hackney:body/1 return the whole response body, and nothing bounds how much the connection process buffers. Any server the client requests can trigger it.
hackney_conn:read_full_body/2 and the HTTP/2 and HTTP/3 accumulators append every chunk until the body ends. With {auto_decompress, true} the buffered body is then inflated in one step with zlib:gunzip/1 or zlib:inflate/2, which return the whole output at once.
This issue affects hackney: from 2.0.0-beta.1 before 4.9.0.4hCVE-2026-115269——
———Allocation of Resources Without Limits or Throttling vulnerability in the HTTP/2 request body reader of mtrudel bandit allows an unauthenticated remote attacker to hold stream processes open indefinitely via WINDOW_UPDATE or empty SETTINGS frames sent while a Plug reads the request body.
Bandit.HTTP2.Stream.do_read_data/4 in lib/bandit/http2/stream.ex waits for each frame with a fresh read_timeout. A stream WINDOW_UPDATE, or the zero-delta window update that every empty SETTINGS frame delivers to all streams, ends the wait without delivering any body data, and the loop starts a new full countdown. A client that sends one of these 9-byte frames before each countdown expires keeps read_body blocked, together with the stream process and whatever the application holds while reading, for as long as it likes.
This issue affects bandit: from 1.4.0 before 1.14.0.4hCVE-2026-114213——
———Allocation of Resources Without Limits or Throttling vulnerability in the WebSocket fragment reassembly of mtrudel bandit allows an unauthenticated remote attacker to exhaust server memory via a fragmented message made of many 1-byte continuation frames.
Bandit.WebSocket.Connection checks max_fragmented_message_size against the payload bytes received, but keeps every non-final continuation frame as a new list cell around its own binary, and nothing limits the number of fragments. A 1-byte fragment costs 7 bytes on the wire but about 150 to 170 bytes of server memory, so one connection can hold about 150 times the configured limit, roughly 1.2 to 1.4 GB at the default 8 MB. Ping frames between fragments keep the connection alive past the idle timeout.
This issue affects bandit: from 1.11.0 before 1.14.0.4hCVE-2026-111101——
———Allocation of Resources Without Limits or Throttling vulnerability in the HTTP/1.1 chunked body reader of mtrudel bandit allows an unauthenticated remote attacker to exhaust server memory and CPU via a request body made of many 1-byte chunks.
The chunked reader in lib/bandit/http1/socket.ex keeps every chunk as its own list cell and binary term until IO.iodata_to_binary/1 joins them, and nothing limits the number of chunks. The :length option of read_body/2 counts only data bytes, so tiny chunks cost far more memory than their payload: 48 MB of 1-byte chunks grew one handler process to about 500 MB and took about 6 s of CPU. The route does not need to read the body, because on keep-alive connections ensure_completed/1 drains unread bodies through the same reader.
This issue affects bandit: from 0.1.0 before 1.14.0.4hCVE-2026-110499——
———Allocation of Resources Without Limits or Throttling vulnerability in the HTTP/2 implementation of mtrudel bandit allows an unauthenticated remote attacker to grow per-connection memory and CPU cost without bound via a large SETTINGS_HEADER_TABLE_SIZE. Bandit resizes its HPACK encoder table to whatever size the client advertises, up to 2^32-1 bytes, and indexes response headers into it. Because responses carry distinct values such as date or content-length, nothing is evicted and the table grows with every response for the life of the connection, while each lookup scans it linearly.
The SETTINGS handling in Bandit.HTTP2.Connection (lib/bandit/http2/connection.ex) passes remote_settings.header_table_size to HPAX.resize/2 with no server-side cap, and encode_headers encodes every response header as :store. Serving N responses on one connection therefore costs O(N²) CPU in the connection process.
This issue affects bandit: from 0.3.6 before 1.14.0.4h