AI Security News: November 20, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. MCP Python SDK Patches an OAuth Flaw That Let a Malicious MCP Server Steer a Client’s Credentials to Its Own Token Endpoint
The Model Context Protocol Python SDK received a high-severity advisory (CVE-2026-115563, CVSS 7.5) for its OAuth client support in mcp.client.auth. The SDK did not validate the authorization server issuer on every discovery path and did not bind stored or pre-provisioned client credentials to the server they belong to, so a malicious or compromised MCP server could name its own token endpoint and receive the client secret, authorization code, PKCE verifier or signed client assertion meant for the real authorization server. Applications are affected when they use the SDK as an HTTP client with an OAuth provider and may connect to an untrusted MCP server while holding legitimate credentials; the fixes are versions 1.30.0 and 2.2.0. Source
2. Jupyter MCP Server Fixes Unauthenticated Management Routes and a Critical DNS Rebinding Path That Hijack Its Jupyter Backend
Jupyter MCP Server (datalayer/jupyter-mcp-server) has two advisories for its standalone streamable-http mode, both fixed in version 1.0.3. CVE-2026-77359 (critical) lets a malicious website use DNS rebinding to send PUT /api/connect to the local server, because the custom management routes lack the Host and Origin checks that protect /mcp. CVE-2026-77318 (high) is the network variant: the same routes do not enforce MCP_TOKEN, so an unauthenticated caller can point the server at attacker-controlled Jupyter infrastructure and reset its context, and DELETE /api/stop is also open. Source
3. k8s-mcp-server Ships With No Authentication on Its Default SSE Transport, Exposing Kubernetes Tools to Any Reachable Client
k8s-mcp-server (reza-gholizade/k8s-mcp-server) is affected by CVE-2026-130007 (high). In its default SSE mode, bound to 0.0.0.0 on port 8080, the tools/call endpoint has no authentication, session token or Origin check, so any client that can reach the port can list and run the server’s Kubernetes tools, including destructive ones, with the cluster privileges the server was configured with. The advisory is fixed in version 1.1.0. Source
4. Pydantic AI Harness Has a High-Severity Denial of Service in Its ANSI-Stripping Tool Output Limit
Pydantic AI Harness is affected by CVE-2026-133161 (high, CVSS 7.5). With strip_ansi=True, the ToolOutputLimits capability (named OverflowingToolOutput before 0.13.0) strips ANSI escape sequences with a pattern that takes quadratic time, so a tool result of a few hundred kilobytes with many unterminated sequences can block the event loop for tens of seconds to minutes. Applications are exposed when an attacker can influence tool output, such as shell commands, file reads, fetched pages or MCP servers; the fix is pydantic-ai-harness 0.70.0, and the default strip_ansi=False is not affected. Source
5. CodeWhale Defaults Automation Tasks to Auto-Approve, Bypassing an Earlier Hardening Fix
CodeWhale, a terminal coding agent, has a high-severity advisory (CVE-2026-83817, CVSS 7.5). The earlier fix for CVE-2026-45374 made the task path require explicit opt-in for auto-approval, but the automation subsystem still resolves an omitted auto_approve to true, which allows silent recurring agent execution without user approval. Versions 0.8.67 and earlier of the codewhale npm package and codewhale-tui crate are affected, and 0.9.0 contains the fix. Source
6. Gortex MCP Server Lets Write Tools Escape Repository Confinement and Reach Code Execution
Gortex (zzet/gortex), an MCP server for code navigation and editing, is affected by CVE-2026-62964 (high, CVSS 7.8). Its write_file, edit_file and move_inline tools exempt absolute paths from repository confinement and skip the symlink guard that the read tools apply, so a single MCP call can write files such as ~/.zshrc or .git/hooks/pre-commit outside every indexed repository and gain code execution on the next shell or commit. Versions before 0.55.0 are affected. Source
7. ToolPulse Shows a Cache Side Channel Can Reconstruct What Tools Claude Code and Codex Run on a Shared Machine
ToolPulse is a paper submitted to arXiv on November 19 that describes a cache-based side-channel attack against tool-using agents. The authors observe that although inference runs remotely, tools execute locally and share system libraries, so Flush+Reload can monitor tool invocations; their Tool Recovery Attack rebuilt tool workflows across 11 LLMs in Codex and Claude Code with an F1 score above 93% at the low end. They also demonstrate task profiling and trigger localization attacks that stay accurate on a server shared by about 20 users. Source
8. Tether Enforces Stateful Security Policies on Coding Agents by Observing Their System Calls in a Sandbox
Tether is an arXiv paper from November 19 proposing runtime enforcement of stateful security policies for agents that run shell commands. Each proposed action runs in an ephemeral sandbox, Tether infers its semantics from the observed system calls, and policies written in a Lean 4 DSL (information flow, access control, network rules) decide whether it may proceed, which covers cases such as a network call made inside a Python script after a sensitive file was read. With one policy shared across all tasks, the authors report blocking 100% of in-scope attacks on coding-agent benchmarks with a benign false-positive rate under 5%. Source