CVEs

## summary `POST /api/chat/completions` accepts an `image_url.url` value that, when it does NOT start with `http://`, `https://`, or `data:image/`, is interpreted as a file id and resolved against the global file table with no ownership check. An authenticated user can therefore set `image_url.url` to another user's file id, the server reads that file from disk, base64-encodes it, and injects the data URI into the LLM request. The user then prompts the LLM to describe / OCR the file and reads t

### Summary `POST /api/v1/calendars/events/{event_id}/update` validates that the caller has **write** access to the calendar the event *currently* belongs to, but does not validate the **destination** `calendar_id` supplied in the request body. The model layer then persists the new `calendar_id` unconditionally. A regular `user`-role account can therefore create an event in their own calendar and immediately move it into any other user's calendar whose ID they know — bypassing the authorizatio

### Summary The spreadsheet-import endpoint `axiosRequestMake` could be used as a generic HTTP proxy. Before the fix it was reachable unauthenticated, and its URL-extension allowlist was a regex tested against the full URL string, so URLs whose query string ended in `.csv` (for example `https://example.com/robots.txt?.csv`) satisfied the gate even though the underlying request was for `robots.txt`. ### Details Three layers of protection now apply to the endpoint: - The controller is decorated

### Summary The base-migration endpoint accepted a caller-supplied URL that the migration worker dereferenced without enforcing protocol or destination, allowing scheme abuse (`file:`, `ftp:`, etc.) and probing of internal HTTP destinations. ### Details The `migrate` endpoint is restricted to the workspace owner role by ACL. The remaining gaps were (a) protocol validation — the controller now parses `body.migrationUrl` as a `URL` and rejects anything whose protocol is not `http:` or `https:` —

### Summary With `NC_SECURE_ATTACHMENTS=true`, an authenticated uploader could deliver `.html` or `.svg` attachments that the browser rendered inline from the NocoDB origin instead of forcing a download. ### Details The signed attachment handler stored response-header overrides under PascalCase keys (`ResponseContentDisposition`, `ResponseContentType`) while the controller that served the file read them under lowercase-hyphen names (`response-content-disposition`). The mismatch dropped the `Con

### Summary A stolen refresh token survived a password-forgot flow and could be used to mint fresh JWTs even after the user reset their password. ### Details `passwordChange` and `passwordReset` deleted the user's refresh tokens, but `passwordForgot` only rotated `token_version` and revoked OAuth tokens — it did not call `UserRefreshToken.deleteAllUserToken(user.id)`. An attacker holding a captured refresh cookie could still exchange it for a new access token after the victim triggered the reco

### Summary The spreadsheet-fetch endpoint (`axiosRequestMake`) accepted URLs whose path contained a permitted extension anywhere in the string, and applied a hand-rolled regex blocklist that omitted `127.0.0.0/8` and `169.254.0.0/16`, allowing the cloud-metadata endpoint to be reached with a crafted URL. ### Details The extension matcher is now anchored to the end of the path or immediately before the query string (`/\.(xls|xlsx|xlsm|ods|ots)(\?|$)/i` and `/\.(csv)(\?|$)/i`), so `http://169.25

### Summary vLLM's `/v1/audio/transcriptions` endpoint limits compressed upload size but not decoded PCM output. A 25MB OPUS file expands to ~14.9GB of float32 PCM at decode time. Tested on vLLM v0.19.0. ### Details `SpeechToTextProcessor` rejects uploads over `VLLM_MAX_AUDIO_CLIP_FILESIZE_MB` (default 25MB) based on compressed byte length, but the audio decoder in `audio.py` accumulates all decoded frames into memory with no size limit before returning: ```python # speech_to_text.py L184-189

# vLLM: incomplete CVE-2026-22778 fix leaks PIL repr addresses via the Anthropic API router **Researcher:** Kai Aizen — SnailSploit (@SnailSploit), Adversarial & Offensive Security Research **Severity:** CVSS 3.1 5.3 (Medium) `AV:N/AC:L/PR:N/UI:N/S:U/C:L/I:N/A:N` **Target:** https://github.com/vllm-project/vllm --- ## Summary The fix for CVE-2026-22778 / GHSA-4r2x-xpjr-7cvv (PRs #31987 and #32319) introduced `sanitize_message` and applied it at four FastAPI exception-handling sites in the O

## Summary Integer truncation of tensor dimensions in vLLM's GGUF dequantize kernels (`csrc/quantization/gguf/gguf_kernel.cu`) causes partial tensor processing. The output tensor is allocated at full size via `torch::empty` (uninitialized memory), but the dequantize CUDA kernel processes only a truncated number of elements. The unfilled portion of the output tensor retains whatever was previously in GPU memory. In multi-tenant inference deployments, this residual GPU memory may contain tensor d

## Summary All temperature validation gates use comparison operators (`<`, `>`), which silently evaluate to `False` for `NaN` and for positive `Infinity` in Python's IEEE 754 float semantics. Both values pass every guard and propagate to GPU sampling kernels, where they produce undefined behavior or CUDA errors that can crash the inference worker. Note: `-Infinity` is correctly caught. ## Root Cause `sampling_params.py:384`: ```python if 0 < self.temperature < _MAX_TEMP: # NaN → False; +Inf

## Summary There is a high severity vulnerability in Traefik's Kubernetes Gateway provider affecting the `crossProviderNamespaces` allowlist. For `HTTPRoute` rules that declare multiple (WRR) backendRefs, Traefik evaluates the allowlist against the target `backendRef.namespace` instead of the route's own namespace. As a result, an `HTTPRoute` created in a namespace that is not allow-listed can reference a cross-provider `TraefikService` such as `api@internal`, `dashboard@internal` or `rest@inte

### Summary The chrome-devtools-mcp daemon writes its PID file with `fs.writeFileSync()` to a deterministic runtime path. On typical macOS environments, and on Linux sessions where `$XDG_RUNTIME_DIR` is unset, that runtime path falls back to `/tmp/chrome-devtools-mcp-<uid>/daemon.pid`. Because the write does not use `O_NOFOLLOW`, a local low-privilege user on the same POSIX host can pre-create `/tmp/chrome-devtools-mcp-<victim_uid>/daemon.pid` as a symlink to a file writable by the victim. Whe

# Pi loads project-local extensions without approval Pi before 0.79.0 loaded project-local configuration and resources from a repository's `.pi` directory without first asking the user to trust that repository. This included project-local extensions, which are executable TypeScript or JavaScript modules loaded into the Pi process. An attacker who controls a repository could place Pi-specific project resources in that repository. If a user then started Pi from that working tree, the project-loc

Hermes WebUI before 0.51.443 contains an authorization bypass vulnerability in the session export endpoint that allows authenticated users to access sessions from other profiles. The _handle_session_export handler in api/routes.py fails to verify active-profile ownership before serializing session data, enabling attackers to exfiltrate foreign session transcripts by guessing or knowing session identifiers.

Hermes WebUI before 0.51.443 contains a broken access control vulnerability in the /api/session endpoint that allows authenticated users to disclose cross-profile session transcripts. Attackers can bypass profile boundary checks by directly querying session IDs belonging to other profiles via GET /api/session?session_id=<foreign_id>&messages=1 to retrieve unauthorized conversation transcripts and metadata.

In Splunk AI Toolkit versions below 5.7.4, a low-privileged user that does not hold the "admin" or "power" Splunk roles could cause the Splunk AI Toolkit to make outbound requests over HTTP to a server that an attacker controls, which could allow for data exfiltration. The vulnerability exists because of an insecure default domain allowlist in the Splunk AI Toolkit, which does not restrict outbound AI agent requests to approved external domains.

A vulnerability in the browser-based version of Cisco Webex App could have allowed an unauthenticated, remote attacker to redirect users to a malicious webpage. Cisco has addressed this vulnerability in the Cisco Webex App, and no customer action is needed. This vulnerability existed due to improper input validation of URL parameters in an HTTP request. Prior to this vulnerability being addressed, an attacker could have exploited this vulnerability by persuading a user to click a crafted URL.