| CVE |
Vendors |
Products |
Updated |
CVSS v3.1 |
| vLLM versions >= 0.6.3 and < 0.9.0 contain multiple regular expression denial of service (ReDoS) vulnerabilities. Several regex patterns — in vllm/lora/utils.py, the phi4mini tool parser, and the OpenAI-compatible serving chat endpoint — are susceptible to catastrophic backtracking. An attacker submitting crafted input with nested or repeated structures can trigger severe CPU consumption and performance degradation, resulting in denial of service. |
| Coturn is a free open source implementation of TURN and STUN Server. Versions prior to 4.10.0 contain a stack buffer overflow in decode_oauth_token_gcm(). A uint16_t nonce_len field read from an attacker-supplied OAuth access token (0-65535) is passed directly to memcpy() as the copy length into a 256-byte stack buffer (oauth_encrypted_block.nonce[256]) without bounds checking. The overflow occurs before AES-GCM authentication is verified, the attacker does not need to know the OAuth key or produce a valid AES-GCM token. Up to 735 bytes of attacker-controlled data are written past the buffer, may corrupt adjacent stack data, including control-flow data depending on compiler, ABI, and mitigations. Requires --oauth mode (non-default). This may provide a plausible RCE primitive depending on exploit mitigations; because coturn is widely deployed for WebRTC TURN/STUN and --oauth is commonly recommended, impact can be broad. This issue has been fixed in version 4.10.0. |
| Microsoft HEIF Image Extensions 1.2.22.0 has an out-of-bounds read because CHEIFItemInfoEntry_GetDataSize can return success while leaving the reported data size as 0. This causes a caller to make a 1-byte allocation. Later, CopyPixels computes copy_size = stride * abs(roi_height) but does not check the source buffer length before a memmove call. |
| picklescan before 0.0.25 fails to detect malicious pickle files that use timeit.timeit() in the __reduce__ method, allowing remote code execution. Attackers can craft pickle files that import dangerous libraries like os and execute arbitrary system commands, which evade picklescan detection and execute when pickle.load() is called. |
| There is an out-of-bounds read vulnerability in the NI grpc-device streaming API due to a missing bounds check that may result in a denial of service. Successful exploitation requires an attacker to supply a specially crafted write request. This affects NI grpc-device 2.17.0 and prior versions. |
| A flaw was found in QEMU. When reading input audio in the virtio-snd device input callback, the `virtio_snd_pcm_in_cb` function did not check whether the iov could fit the data buffer, potentially leading to a heap out-of-bounds write. This issue exists due to an incomplete fix for CVE-2024-7730. |
| An integer overflow vulnerability was found in the virtio-snd device via PCM_INFO requests from the guest. A malicious guest can provide out-of-bounds stream counts, potentially leading to unbounded memory allocation on the host and a denial of service condition. |
| OpenEXR is the reference implementation and specification for the EXR image format, widely used in the motion picture industry. In versions 3.4.0 through 3.4.11, an integer overflow in ht_undo_impl() in src/lib/OpenEXRCore/internal_ht.cpp leads to a heap-buffer overflow when decoding a crafted HTJ2K-compressed EXR file. decode->channels[i].width (int32_t) is multiplied by bytes_per_element in 32-bit signed arithmetic. With large widths (e.g., >= 536870912 for FLOAT data), this overflows, producing a corrupted offset that is later used for pointer arithmetic and can cause a heap out-of-bounds write. The same unchecked multiplication pattern appears in two other HTJ2K paths (bytes-per-line accumulation and pixel-line pointer advancement). As with related CVE-2026-34378 through CVE-2026-34589 fixes in other codecs, validating only after the multiplication is too late because the value may already be overflowed. This issue has been fixed in version 3.4.12. |
| Crypt::OpenSSL::PKCS12 versions before 1.96 for Perl permits a heap OOB read in print_attribute UTF8STRING path.
print_attribute() copies a UTF8STRING ASN.1 attribute value into a heap buffer sized exactly to its declared length via strncpy, leaving no NUL terminator. Downstream callers run strlen() on the result and pass the inflated length to newSVpvn(), copying attacker-influenced adjacent heap bytes into a Perl scalar. |
| deepstream is a server that allows clients and backend services to sync data, send messages and make rpcs at scale. Versions prior to 10.0.5 are vulnerable to Prototype Pollution. Exploitation can lead to potential privilege escalation from any authenticated user with write permission to any record. This issue has been fixed in version 10.0.5. |
| @microsoft/kiota-http-fetchlibrary provides TypeScript libraries for Kiota-generated API clients. In versions 1.0.0-preview.97 through 1.0.0-preview.101, `@microsoft/kiota-http-fetchlibrary`'s `RedirectHandler` is documented as stripping `Authorization` and `Cookie` from cross-origin redirect targets, but the default `scrubSensitiveHeaders` callback in `RedirectHandlerOptions` uses case-sensitive property deletion (`delete headers.Authorization`, `delete headers.Cookie`) on a headers object that `FetchRequestAdapter.getRequestFromRequestInformation` has already lower-cased. The delete therefore targets keys that do not exist, the scrub is a no-op, and any Bearer token or Cookie attached by a kiota-generated SDK is forwarded to an attacker-controlled host across a 30x redirect. This is reachable in the default middleware chain (`MiddlewareFactory.getDefaultMiddlewares`) with no custom configuration, and applies to every kiota-generated TypeScript SDK that uses `BaseBearerTokenAuthenticationProvider` or any other authentication provider that sets the `Authorization` request header. Version 1.0.0-preview.102 patches the issue. |
| Inappropriate implementation in Passwords in Google Chrome on Android prior to 149.0.7827.115 allowed a remote attacker who had compromised the renderer process to bypass site isolation via a crafted HTML page. (Chromium security severity: High) |
| radvd is a router advertisement daemon for IPv6. Prior to version 2.21, the `radvdump` utility shipped with radvd contains a stack buffer overflow in the Route Information option parser. When processing a crafted ICMPv6 Router Advertisement, `print_ff()` copies up to 2032 bytes from attacker-controlled packet data into a 16-byte `struct in6_addr` on the stack, overflowing by up to 2016 bytes. Note that the main `radvd` daemon is not affected by the vulnerability. Version 2.21 patches the issue. |
| A vulnerability was identified in BerriAI litellm up to 1.82.2. This affects an unknown function of the file litellm/proxy/auth/user_api_key_auth.py of the component M2M JWT Handler. Such manipulation leads to improper authorization. The attack can be launched remotely. A high complexity level is associated with this attack. The exploitability is reported as difficult. The exploit is publicly available and might be used. The vendor was contacted early about this disclosure. |
| ImageMagick before 7.1.2-15 (and 6.x before 6.9.13-40) contains a heap out-of-bounds read in the PCD coder's DecodeImage loop. A crafted PCD file can trigger a one-byte heap out-of-bounds read during image decoding, resulting in denial of service and potential disclosure of an adjacent heap byte. |
| Untrusted user data was passed verbatim to Excel exports for administrators. This allowed formula injection which can be used to compromise the environment of the user loading the file or other data in the file. |
| JTL Shop versions 5.2.0 through 5.7.1 contains a server-side template injection vulnerability that allows unauthenticated attackers to inject malicious template syntax due to unsanitized user-supplied input passed to the Smarty template engine. Attackers can exploit this flaw to read sensitive server-side values such as database credentials and encryption keys, and on versions 5.4.0 through 5.7.1, leverage registered Smarty modifiers including unserialize and file_get_contents to write a webshell to the web root and execute arbitrary commands as the web server user. |
| Improper escaping of database table names in the CaptureChangeMySQL Processor included with Apache NiFi 1.2.0 through 2.9.0 allows for injecting SQL commands using crafted naming. Manual quoted boundaries added in Apache NiFi 1.8.0 narrowed the scope of potential injection options, but did not cover additional strategies. Apache NiFi installations that do not use the CaptureChangeMySQL Processor are not subject to this vulnerability. Upgrading to Apache NiFi 2.10.0 is the recommended mitigation, which incorporates more robust identifier escaping. |
| ImageMagick before 7.1.2-15 and 6.9.x before 6.9.13-40 contains an integer overflow in the PSB (PSD v2) RLE decoding path (ReadPSDChannelRLE in coders/psd.c) that causes a heap out-of-bounds read on 32-bit builds. Processing a crafted PSB file can lead to information disclosure or a crash. |
| Backpropagate is a Python library for fine-tuning large language models on a single GPU. In versions 1.1.0 and 1.1.1, the optional Reflex web UI exposes a training control plane without authentication: dataset upload, model load, training start/stop, multi-run orchestration, GGUF export, and HuggingFace Hub push. The CLI accepts two operator-facing flags intended as security controls: --auth user:pass — documented as "require HTTP Basic authentication on every request to the UI." and--share — documented as "expose the UI on a public address; requires --auth." When --auth user:pass is passed, the CLI prints Auth: enabled (user: <username>) to confirm to the operator that authentication is active, then exports BACKPROPAGATE_UI_AUTH=user:pass to the subprocess that launches the Reflex backend. The Reflex backend (backpropagate/ui_app/**) never reads BACKPROPAGATE_UI_AUTH. No authentication middleware is registered. No request-level guard runs. No WebSocket upgrade guard runs. Any client that reaches the bound port — local or remote, depending on whether --share is used — has full UI access. An inline comment at backpropagate/cli.py:1217-1218 in the v1.1.0 source documents the gap: "For Phase 1 the variable is exported but Reflex doesn't read it yet." This comment was internal-facing; the user-facing documentation (README, CHANGELOG, SHIP_GATE) advertised the contract as enforced. An attacker who reaches the bound port can read uploaded datasets, trigger arbitrary training runs against any local base models as well as read their paths, trigger HuggingFace Hub pushes and cause disk-fill DoS. This issue has been fixed in version 1.2.0. If developers cannot immediately upgrade to 1.2.0 run backprop ui with no flags so it binds to localhost, use SSH port-forwarding (ssh -L 7860:localhost:7860 <training-host>) instead of --share for remote access, and audit any host previously launched with --share, re-issuing any HF tokens used during those sessions. |
