Decap CMS versions prior to 3.11.0 have a critical path traversal vulnerability in the local proxy containment guard. The vulnerability arises from improper pathname limitation using plain string prefix comparison without validating path separators. This flaw allows attackers to access sibling directories with names starting like the repository directory, enabling unauthorized reading, writing, or deletion of files outside the intended repository root. The issue is fixed in version 3.11.0.
vLLM is an inference and serving engine for large language models. Prior to 0.24.0, the input_audio handling path for /v1/chat/completions calls AudioMediaIO.load_bytes or AudioMediaIO.load_file without passing VLLM_MAX_AUDIO_DECODE_DURATION_S to the shared audio decoder. An unauthenticated client can therefore submit a small compressed audio input that expands into a very large float32 PCM allocation, bypassing the duration guard already used by /v1/audio/transcriptions and causing an out-of-memory worker crash. Inline data URLs reach this path without being bounded by VLLM_AUDIO_FETCH_TIMEOUT. The issue affects deployments serving an audio-capable model, and authentication changes only the deployment-specific reachability. This issue is fixed in version 0.24.0.
MCP Context Forge is an AI gateway, registry, and proxy for MCP, A2A, REST, and gRPC APIs. Prior to 1.0.2, the python_sandbox_server in mcp-servers/python/python_sandbox_server/src/python_sandbox_server/server_fastmcp.py exposes raw getattr through safe_builtins, omits a required _getattr_ guard, and relies on validate_code checks for literal dangerous dunder strings. An attacker can construct dunder names at runtime, traverse the Python class hierarchy, reach subprocess.Popen, and execute OS commands with the server process privileges through the execute_code MCP tool. The HTTP/SSE transport can expose this tool without authentication, while stdio-only deployments have reduced network reachability. The issue affects the python_sandbox_server subproject and does not directly affect the core Context Forge gateway or proxy components. This issue is fixed in version 1.0.2.
### Summary
The 50-megapixel decode guard exists only on the task-attachment preview path. Avatar and project-background uploads decode uploaded images with no pixel cap. Worse, the avatar resize fixes the output height at 1024 and derives the width from the aspect ratio, so a tiny extreme-aspect-ratio PNG expands to an enormous output image — an input-side pixel cap would not catch it.
### Details
The only `maxPixels` check (50MP) is in `TaskAttachment.GetPreview` (`pkg/models/task_attachment.go` ~lines 332-338). No such check guards:
- avatar upload: `pkg/modules/avatar/upload/upload.go` (~lines 82, 137, 141)
- project background: `pkg/modules/background/handler/background.go` (~lines 174, 269, 289)
`imaging.Resize(img, 0, 1024, imaging.Lanczos)` (upload.go ~line 141) fixes height=1024 and derives width from the aspect ratio: a 20000x10 input yields a ~2,048,000 x 1024 output (~2.1 billion pixels), so a few-hundred-byte file drives huge CPU and memory.
### PoC (verified at runtime against v2.5.0)
```
PUT /api/v1/user/settings/avatar/upload avatar=8000x8000 PNG (64MP, 192KB) -> 200 (accepted; exceeds the 50MP attachment cap)
PUT /api/v1/user/settings/avatar/upload avatar=20000x10 PNG (681 bytes) -> 200 after ~19.6s of server processing
```
Contrast (guard present): uploading the 64MP PNG as a task attachment and requesting its preview returns in ~1ms without decoding — `GetPreview` rejects it via `maxPixels` and falls back to the raw file.
### Impact
A small crafted upload drives disproportionate CPU and memory on the avatar and background paths. Repeated or concurrent requests can exhaust server resources. Amplification comes from both the missing input pixel cap and the height-fixed resize, so an input-side cap alone is insufficient.
### Fix
Apply a pixel-dimension cap (as on the attachment path) to the avatar and background decode paths, and bound the resize output dimensions (cap width as well as height).
Joomla Extension - joomshaper.com - Missing Directory Confinement in Media Rename Allowing Arbitrary File Rename in SP Page Builder (Free and Pro) 4.0.0 - 6.9.0 - The media rename task applied neither of the directory boundary checks used by the folder operations in the same controller, and its validation guard required only that either a media record exist for the supplied identifier or that the supplied path be present in #__spmedia, rather than both. The identifier and the path were consequently never checked against one another, so any valid media identifier could be paired with an unrelated filesystem path, and the STR input filter left traversal sequences intact. An attacker could rename files elsewhere in the installation, including renaming configuration.php to take the site offline.
A security flaw has been discovered in GNU libredwg 0.13.4. This impacts the function DWG_TABLE of the file src/dwg.spec of the component Layer Encoding. Performing a manipulation results in null pointer dereference. The attack needs to be approached locally. The exploit has been released to the public and may be used for attacks. Upgrading to version 0.14 will fix this issue. The patch is named f5b548c4c1697d66c3dabd0f6a49280a14365a3a. The affected component should be upgraded. The FIELD_HANDLE macro itself is NULL-safe (emits null_handle) - only the two raw zeroing assignments added by 27118c40 ("encode: also disable LAYER.material") dereferenced a NULL material handle; the fix restores the file's existing if (_obj->style) guard convention for material.
A vulnerability in the Linux kernel's svcrdma component allowed Read lists with cumulative segment lengths exceeding the page budget to cause a NULL pointer dereference during DMA mapping. This occurs because the runtime guard can be bypassed when a segment exactly fills the page array, leading to an out-of-bounds access. The issue has been resolved by validating and rejecting messages that would overflow the page budget at decode time.
In the Linux kernel, the following vulnerability has been resolved:
RDMA/uverbs: Guard legacy bundles without method_elm
The legacy write() path dispatches through a uverbs_api_write_method, but
the uverbs_attr_bundle passed to provider code does not have an ioctl
method element. If malformed provider input causes the common uverbs
validation code to emit an error message, uverbs_get_handler_fn()
dereferences the uninitialized method_elm pointer.
Initialize method_elm explicitly for legacy bundles and make
uverbs_get_handler_fn() return NULL when no ioctl method is present. The
legacy dispatcher continues to use its local write method, while the ioctl
path continues to use the registered ioctl handler.
A vulnerability in the Linux kernel's svcrdma component allows oversized Read segments to be processed without proper validation, potentially causing resource leaks. The issue arises because the RPC/RDMA Read list decoder does not cap segment lengths, leading to large dynamic allocations that may exceed the receive context's page budget. This can trigger a page-overrun guard, which leaks the resource by not returning the acquired read-write context. The vulnerability has been addressed by rejecting segments that exceed the page budget during decoding and ensuring proper resource cleanup on overrun detection.
AVideo through commit c3edcc274c389816d434acadac07ee78eaf330c1 contains an open redirect in objects/playlistSort.php. Because the endpoint is not a *.json.php script, AVideo's automatic CSRF guard (autoCSRFGuard()/forbidIfIsUntrustedRequest()) does not run, and when the request includes the sort parameter the script issues a Location header set to the unvalidated $_SERVER['HTTP_REFERER'] value without calling isSafeRedirectURL(). A remote unauthenticated attacker can therefore induce a logged-in user who can manage the targeted playlist to submit a cross-origin POST with a crafted Referer, causing the victim's playlist to be reordered and the victim's browser to be redirected from a trusted AVideo URL to an attacker-controlled site for phishing. No patched version is available.
A security researcher discovered an authentication bypass vulnerability in a YCombinator startup's platform that allowed unauthorized access to patient information, including personally identifiable information (PII) and protected health information (PHI). The vulnerability exposed API keys and tenant data behind an authentication guard. Attempts to responsibly disclose the issue to the startup's CEO were met with dismissal and blocking of the researcher. No patch or remediation information is provided.
OpenPanel analytics platform contains a server-side request forgery (SSRF) vulnerability in its data importer component. The importer fetches a caller-supplied URL using a plain fetch method without applying the existing SSRF guard. Authenticated organization members can cause the server to connect to arbitrary internal or external addresses. The HTTP response status and text are stored and returned to the user, enabling internal network scanning. Additionally, if the internal response is in a specific CSV format, it can be ingested as analytics events visible to the attacker. No patched version is available at the time of publication.
GitPython versions up to 3.1.58 contain a vulnerability in the TagReference.create() method where a positional argument 'reference' can bypass the unsafe option guard for the '--file' parameter. This allows an attacker who can control the 'reference' argument to cause an arbitrary local file read, with the file contents returned as the annotated tag message. The issue is an incomplete fix of a previous vulnerability and affects applications that forward user-controlled values into this method. A fix is available in version 1.4.0_3.
The Hustle WordPress plugin before 7.8.14.2 does not prevent shortcodes in submitted form values from being executed when it substitutes those values into the message it returns after submission, because the guard it applies can be defeated by nesting, allowing unauthenticated users to run any shortcode registered on the site.
The Custom Contact Forms plugin for WordPress is vulnerable to authorization bypass in all versions up to, and including, 7.16. This is due to the plugin not properly verifying that a user is authorized to perform an action. This makes it possible for authenticated attackers, with contributor-level access and above, to permanently force-delete arbitrary posts of any post type (including pages, administrator-authored posts, and WooCommerce products) and write arbitrary ccf_field_* post meta onto any post regardless of ownership or post type. The top-level form ID is checked via edit_post/publish_posts, but the nested fields[].ID and choices[].ID paths processed by _create_and_map_fields() and _create_and_map_choices() carry no equivalent capability or post-type guard, leaving those sinks fully exposed while delete_item() and delete_submission() contain explicit post-type restriction fixes demonstrating the developer's awareness of scoping requirements.
In the Linux kernel, the following vulnerability has been resolved: HID: nintendo: fix out-of-bounds read in joycon_ctlr_read_handler() joycon_ctlr_read_handler() casts an incoming HID input report to struct joycon_input_report and parses it, guarding the cast only with a 12-byte length check: if (size >= 12) /* make sure it contains the input report */ joycon_parse_report(ctlr, (struct joycon_input_report *)data); struct joycon_input_report is 49 bytes: a 13-byte header followed by a union whose IMU arm is 36 bytes. For an IMU report joycon_parse_report() -> joycon_parse_imu_report() walks that union (struct offsets 13..48), so a report of exactly 12 bytes with data[0] == JC_INPUT_IMU_DATA passes the guard yet is read up to 37 bytes past its declared length. The over-read bytes are decoded into accelerometer/gyroscope values and forwarded to userspace through the "(IMU)" input device, leaking driver-internal memory. data[0] and size are fully controlled by a malicious or spoofed Joy-Con/Pro Controller. Receive buffers are sized to the maximum report length, so this is an over-read within the allocation rather than a slab OOB, but the decoded bytes still reach userspace. The sibling subcmd path in joycon_ctlr_handle_event() already bounds the same cast correctly: if (size < sizeof(struct joycon_input_report) || data[0] != JC_INPUT_SUBCMD_REPLY) break; Use the same sizeof(struct joycon_input_report) bound here.
In the Linux kernel, the following vulnerability has been resolved: HID: core: fix OOB read of field->usage in hid_set_field() hid_set_field() hands field->usage + offset to hid_dump_input() before the guard that bounds offset: hid_dump_input(field->report->device, field->usage + offset, value); if (offset >= field->report_count) { hid_err(...); return -1; } Under CONFIG_DEBUG_FS hid_dump_input() dereferences that pointer, with buf = hid_resolv_usage(usage->hid, NULL). The usage[] array is allocated inline with the hid_field in hid_register_field() and holds field->maxusage entries, so an offset past it reads off the end of the kvzalloc()ed allocation and into a neighbouring object. Had the guard run first, offset < report_count <= maxusage would already have confined the pointer to the array. A caller supplies such an offset today. picolcd_fb_send_tile() validates only report->maxfield before issuing hid_set_field(report->field[0], 11 + i, ...) for i = 0..31, so its offsets are fixed at 11..42 and are never checked against the bound field. When the device registers that field with fewer usages, the framebuffer deferred-io work drives the read on every tile. KASAN reports a 4-byte slab-out-of-bounds read in hid_dump_input() below hid_set_field(), and the same boot logs "offset (1) exceeds report_count (1)" from the guard that runs only afterwards. Move the hid_dump_input() call below the guard. Because field->maxusage >= field->report_count, the guard then establishes that field->usage + offset lies inside the array before it is dereferenced, for every caller and without changing behaviour on the valid path. Discovered by XBOW, triaged by Baul Lee <baul.lee@xbow.com>
In the Linux kernel, the following vulnerability has been resolved: HID: magicmouse: prevent unbounded recursion in magicmouse_raw_event() magicmouse_raw_event() handles DOUBLE_REPORT_ID (0xf7) packets, which pack two touch reports into one, by splitting the packet and calling itself on each half. The only guard against runaway recursion is a "size < 1" check, which stops zero-sized calls but does not bound the recursion depth. A malicious HID device that matches this driver can send a report starting with DOUBLE_REPORT_ID and filled with the sequence [0xf7, 0x00]. Each level consumes two bytes and recurses on the remainder, so an incoming report of up to HID_MAX_BUFFER_SIZE (16 KiB) drives roughly 8000 nested calls. That easily exhausts the 16 KiB kernel stack, leading to a stack overflow: a panic with CONFIG_VMAP_STACK, or memory corruption without it. A double report only ever wraps two normal reports; it is never legitimately nested. Refuse to re-enter the DOUBLE_REPORT_ID case from a recursive call so the recursion depth is bounded to two, while all valid packets keep being parsed exactly as before.