AI Security News: November 21, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. FlowIntel Fixes Chatbot and AI Report Flaws That Let an Injected Model Reach Administrative API Routes and Mix Up User Credentials
FlowIntel, an open-source case management platform with an AI chatbot and report generator, has two advisories published on November 20. CVE-2026-133631 (high) covers several flaws fixed in version 4.1.0: the chatbot API tool let the language model call any FlowIntel endpoint, including administrative routes and state-changing GET operations, with no allow-list, so an attacker who can influence the model’s instructions through untrusted case content could change data or reach restricted endpoints. Note variable references were resolved with the viewer’s privileges, which could pull private case data into AI report requests, exports and MISP sends, and report generation passed the Ollama bearer token and full case content as command-line arguments visible to local users. CVE-2026-133853 (medium) is a separate flaw in which the chatbot used one shared MCP session and a process-wide FLOWINTEL_API_KEY, so one user’s credentials could be applied to another user’s tool call. Source
2. llama.cpp Patches a High-Severity Integer Overflow in Its hrm_text Model Loader That Crafted GGUF Files Can Use to Corrupt the Heap
llama.cpp is affected by CVE-2026-133710 (high, CVSS 7.8). In builds b11003 up to before b12556, load_arch_hparams() in the hrm_text loader can be bypassed through crafted GGUF metadata, because layer factors wrap modulo 2^32 and bypass the block count check. load_arch_tensors() then writes tensor pointers past model.layers, which corrupts heap memory and can lead to code execution when a user loads an attacker-supplied model. The fix is in build b12556. Source
3. Flyto2 Core Follows Server-Supplied Pagination URLs Without Rechecking Them, Opening a Second-Order SSRF (CVE-2026-77315)
Flyto2 Core (flytohub/flyto-core), an Apache-2.0 workflow automation engine that describes itself as showing proof of what AI agents did, has a high-severity advisory (CVSS 8.5). Its http.paginate module validates only the initial base_url, then follows Link: rel="next" headers and cursor URLs from the server response without revalidating them. A malicious first-page response can therefore point at an internal address or cloud metadata endpoint, and the internal response body is aggregated into the items returned to the caller. Versions up to 2.26.7 are affected, and 2.26.8 contains the fix. Source
4. NVIDIA TensorRT-LLM Has a Race Condition on Concurrent Requests That May Lead to Data Tampering (CVE-2026-24205)
NVIDIA TensorRT-LLM is affected by a medium-severity flaw (CVSS 6.4) on all platforms, in which a user can trigger a race condition by running concurrent requests. According to the advisory, a successful exploit may lead to data tampering. The GitHub advisory, published on November 20, gives no affected version range in its summary, so operators are directed to NVIDIA’s security bulletin for the patched release. Source
5. llamafile Has a Remotely Reachable Out-of-Bounds Read in Its Zip Container Parser, Fixed in Version 0.10.6 (CVE-2026-133498)
llamafile, Mozilla AI’s single-file LLM distribution, is affected by a medium-severity flaw (CVSS 6.3) in llamafile_open_zip in llamafile/llamafile.c. Manipulating the size argument in the container parser causes an out-of-bounds read, and the advisory states the attack can be carried out remotely. Versions up to 0.10.5 are affected, and upgrading to 0.10.6 fixes it. Source