Five score 9.3, and NVD has scored none
NVD published 16 VulnCheck-sourced CVE records between 01:17:09 and 01:17:20 UTC today, and 15 of them cover ten LLM agent platforms that teams run on their own servers. Four of the five 9.3s: piflow 2.0.0 through 2.0.1, whose POST /chat and /chat/stream need no authentication and whose exec_shell tool runs subprocess.run with no human approval or allow-list; Langchain-Chatchat through 0.3.1.3, whose POST /tools/call invokes any registered tool by name, unauthenticated; DB-GPT through 0.8.2, whose /api/v1/chat/react-agent endpoint relies on a mock get_user_from_headers that grants every caller the admin role; and Letta through 0.16.8, whose shipped compose.yaml does not enable LETTA_SERVER_SECURE, so the password middleware is never added.
Every score on the 16 is VulnCheck's own. NVD has assigned none of its own and has marked the two DB-GPT records Deferred. The other ten records run from 8.7 down to 6.9. Six are more ungated command or code execution — OpenSRE, whose execution_allowed() always returns True; EvoScientist 0.1.0 before 0.2.5, delegating commands to async sub-agents with no interrupt_on gate; Coze Studio before 0.5.0, whose workflow code nodes run in an unsandboxed python3 subprocess for any registered user. Four are other defects: tenant isolation at LangAlpha, an IDOR at Coze Studio, header impersonation and secret exposure at Letta. Three turn on a shipped deployment file, so check the compose file as well as the version.
Meta's new terms put AI mistakes on the customer
Meta's revised commercial terms are scheduled to take effect on 30 October and could still change. As drafted, they say AI features "may make mistakes" and may not always produce "accurate, complete, or suitable results", and that the customer is solely responsible for, and assumes any and all risks associated with, those features and any resulting content, CIO reported on 19 October. The customer is also responsible for anything done by a partner or contractor. Using a Meta product grants Meta a licence to content covered by intellectual-property rights, including material posted on the customer's behalf by an agency.
Meta removed the reference to training its models on enterprise data and now cites developing and maintaining advertising and business products. The draft also let Meta use "web scrapers, bots, robots, spiders, crawlers" against content associated with a customer's products. A Meta public affairs manager told CIO that they apply only to its ads and business products, that crawler use is unchanged, that robots.txt still signals content owners' preferences, and that Meta had clarified this in the terms, though the proposed wording had yet to be altered. Infotech's Thomas Randall called the terms "a one-way street". A buyer with a Meta business account should check who carries the risk for outputs.
Quick Hits
- DeepSeek's checkpoint now has an independent score. Artificial Analysis puts V4.1 Flash 1017 at Intelligence Index 45 and #2 of 121 open-weight models of its size class, against 39 and #10 for the September release.
- Germany's economy ministry signed a memorandum of understanding with Nvidia. The government released no details, Reuters reported on 19 October. Merz told business leaders in Berlin: "Please open your doors for AI to make use of your data."
- A short request ID aborts other requests on SGLang servers through 0.5.22. CVE-2026-112266 says single-hex-character rids on streaming /generate, then a disconnect, trigger abort logic terminating other clients' requests. VulnCheck scores it 8.2 on CVSS 4.0, 5.9 on 3.1.
- MLflow's FileStore backend writes where the trace ID points. CVE-2026-53995 (8.8 on CVSS 4.0, VulnCheck's) says an unauthenticated crafted trace ID at POST /api/3.0/mlflow/traces escapes the experiment directory in 3.17.1 and earlier. FileStore was default before 3.7.0, opt-in after.