Apache OpenNLP's built-in regex name finder for latitude and longitude coordinates uses a pattern vulnerable to excessive backtracking. Text supplied to that finder can trigger a denial of service; opennlp-tools 2.0.0 before 2.5.13 and opennlp-runtime 3.0.0-M1 before 3.0.0 are affected.
A memory corruption vulnerability dubbed Java4Shell in the OpenJDK runtime was discovered using AI, according to a single report. The flaw could let unauthenticated attackers take over affected systems, threatening enterprises that rely on Java applications.
RunAnywhere open-sourced TrnRT, an inference engine for AWS Trainium chips, after three weeks of development. It serves two instances of GLM-5.3 Flash (FP8) on a single trn2.48xlarge box with a cache-aware router, reporting 3.57M tokens per minute on one benchmark.
Linx Security announced four capabilities extending identity governance to AI agents, including Shadow AI Agent Detection for unsanctioned agents and Agent Runtime Control that evaluates agent actions inside existing gateways. It also introduced Autopilots for building governed agents and Prompt-to-Workflow, which converts plain-English descriptions into tested agentic workflows.
Airlock Digital released version 7.1, adding Agentic Control with Agentic Steering capabilities that evaluate AI agent requested behavior at runtime. The feature extends Deny by Default Application Control to block unauthorized agent actions before execution.
Vim before 9.2.1346 contains a command-injection flaw in runtime/ftplugin/ps1.vim, where the PowerShell keyword-lookup (keywordprg) setting fails to properly neutralize special elements. The issue is rated low severity under CWE-78, builds on an incomplete fix from GHSA-r5v6-q6j8-8qw2, and no CVE has yet been assigned.
Vim's CSS omni-completion script (runtime/autoload/csscomplete.vim) passes the filename in an @import rule to glob() without escaping, allowing OS command injection (CWE-78). The flaw affects Vim before 9.2.1342 and is rated low severity, with a CVE requested but not yet assigned.
Obsidian Security announced a Runtime Policy Decision Engine that lets security teams allow, block, or require human approval for AI agent actions before they execute. The engine uses context about the agent's operator, its access, and involved third-party applications, building on Obsidian's existing AI agent discovery and governance coverage.
Siemens published security advisory AV26-1139 covering vulnerabilities in Mendix Studio Pro 11 and 11.12, ROS#, SIMATIC BATCH V9.1, SIMATIC Information Server, SIMATIC PCS 7 V9.1, SIMATIC Process Historian, and SIMATIC WinCC Runtime Professional V17 through V21. Fixed releases exist for Mendix and ROS#, while several SIMATIC products are affected in all versions.
SAP's November 2026 monthly security rollup (AV26-1137) lists vulnerabilities in NetWeaver AS ABAP Internet Communication Manager, Web Dispatcher, and Gateway across kernel versions including 7.53 through 9.20. The advisory does not detail individual vulnerability types or severity, and affected organizations should apply the corresponding SAP Security Notes.
Sysdig has released Sysdig AI Defense, which monitors and controls AI agent activity across endpoints and cloud environments, allowing, holding or blocking agent actions against company policies. It also logs agent activity and policy decisions for investigations and audits.
Sysdig has added runtime security and governance features for AI agents to its platform. The platform uses runtime visibility to track agent actions and block risky behavior before it reaches production systems.
The article argues that agent runtime security products can stop bad agent actions in real time only for agents whose traffic passes through an enterprise-owned gateway. It says most agents never touch such a gateway and cites an incident in which an employee asked an AI assistant to trigger a mass-email automation.
Microsoft Security Blog research details an intrusion chain starting with help desk impersonation to gain remote access, followed by portable runtime staging and Active Directory Certificate Services assessment. It analyzes ESC1 misconfigured certificate templates and ESC8 NTLM relay abuse used to escalate privileges and obtain Kerberos tickets.