What's new in agentic AI this week: standards, security patches and observability (10–17 November 2026)
By Jarvis Nelson Osorio · November 17, 2026
Jarvis Nelson Osorio (also Jarvis Osorio) is a tech entrepreneur and AI innovator based in Miami Beach, FL, and the builder of Ambient Memory, mcp-pordego, agentscrub and Agent Replay Lab. This is a sourced roundup of what actually shipped or was disclosed in agentic AI between 10 and 17 November 2026, with links to the primary documents.

Standards: MCP and Agent Skills under vendor-neutral governance On 11 November 2026, the Google Open Source Blog published a conversation with Kurtis Van Gent of Google Cloud, a core maintainer of both the Model Context Protocol (MCP) and Agent Skills, and Daryl Ducharme of the Google Open Source Programs Office: “Scaling AI agents with open standards: Kurtis Van Gent on MCP and Agent Skills joining the AAIF,” at https://opensource.googleblog.com/2026/11/scaling-ai-agents-open-standards-mcp-aaif.html.
The post describes two complementary open standards. MCP connects agents to live tools and data. Agent Skills load procedural instructions and scripts on demand. Following Anthropic's donation of the Agent Skills specification to the Linux Foundation's Agentic AI Foundation (AAIF) alongside MCP, engineers from Google, Anthropic, OpenAI and the wider ecosystem are now co-maintaining the standard under vendor-neutral governance.
Van Gent argues that MCP commoditized tools, and the unintended consequence was tool and context bloat. Agent Skills address that with progressive disclosure: load only a lightweight YAML frontmatter name and description at startup, and pull the full SKILL.md instructions, reference files or scripts into context only when a task requires them. Skills teach an agent how to approach a problem; MCP is the live execution interface. The post also points to work on Agent Plugins and on serving Agent Skills over MCP, and says what comes next is packaging, discovery, naming-conflict resolution and versioning. Portable-skill resources: https://agentskills.io https://github.com/google/skills and https://github.com/agentskills

Security: a week of MCP advisories The same week produced a tight cluster of official disclosures around MCP servers, developer workbenches and gateways. I cover the technical detail in a companion post; the headlines belong here.
On 11 November 2026, the Model Context Protocol servers repository published GHSA-6hqx-gqrx-c5r2, a high-severity advisory for mcp-server-git. The reference Git MCP server ran Git commands inside client-chosen repositories without neutralizing the repository's own configuration. Git honors command-bearing keys such as core.fsmonitor, so a call to the read-only git_status tool on an untrusted repository could execute an attacker-chosen program with the MCP server's privileges. The patched release is mcp-server-git 2026.11.11. Advisory: https://github.com/modelcontextprotocol/servers/security/advisories/GHSA-6hqx-gqrx-c5r2
On 12 November 2026, the Inspector project published GHSA-9w48-x22j-rrm7. A connected MCP server that answers a JSON-RPC request with an unbounded text/event-stream POST response could make the Inspector web backend run out of memory and drop every session on that backend. Affected versions are 2.0.0 through 2.14.0; the patch is 2.15.0. Advisory: https://github.com/modelcontextprotocol/inspector/security/advisories/GHSA-9w48-x22j-rrm7
The same day, IBM published a security bulletin for ContextForge MCP Gateway versions 1.0.0 through 1.1.0, tracked as CVE-2026-110278. When identity propagation is enabled, a per-server setting that should turn it off is ignored when invoking tools and reading resources. IBM's CVSS base score is 7.4. The fix is 1.1.1; a workaround is to disable identity propagation globally. Bulletin: https://www.ibm.com/support/pages/security-bulletin-ibm-contextforge-mcp-gateway-affected-information-exposure-vulnerability
On 16 November 2026, AWS Labs published GHSA-mqm3-w68j-75mw for the AWS Pricing MCP Server. Before version 1.0.31, the cost-report generation tool could, under certain circumstances, write attacker-influenced content outside the intended output directory. The tool had been annotated as read-only, so auto-approve clients could run it without confirmation. No managed AWS service is affected. The patch is 1.0.31, and readOnlyHint is now False. Advisory: https://github.com/awslabs/mcp/security/advisories/GHSA-mqm3-w68j-75mw
The pattern is consistent with what mcp-pordego and agentscrub were built to inspect: tool-level trust is still being granted too early, and “read-only” is not a safety property.
Frameworks and enterprise observability Microsoft released Agent Framework Python 1.26.0 this week. The GitHub release is dated 13 November 2026 (https://github.com/microsoft/agent-framework/releases/tag/python-1.26.0); the project changelog dates the 1.26.0 section 12 November 2026 (https://github.com/microsoft/agent-framework/blob/main/python/CHANGELOG.md). Verified additions include loop_return_final_only on create_harness_agent(); an optional experimental SecureAgentConfig.message_labeler callback; structured-output support in the GitHub Copilot integration through response_format; and a default 30-second Hyperlight execution budget.
Separately, on 10 November 2026, Microsoft published “Microsoft Ignite 2026: Your Guide to Agent 365,” ahead of Ignite (17–20 November 2026). The guide says Agent 365 is featured as the observability layer for agents, models and assistants in the opening keynote (KEY01, 17 November, 10:00 AM PT), with a follow-on session, INV642, on 18 November. A lightning talk on 19 November, LTG1238, covers managing MCP servers, plugins and skills for AI agents. Guide: https://techcommunity.microsoft.com/blog/agent-365-blog/microsoft-ignite-2026-your-guide-to-agent-365/4571750

Also worth noting, with dates attached Two research posts landed just before this seven-day window and were still circulating this week, so I will date them honestly. On 6 November 2026, Google published “Agent Distraction” (InfoQ covered it on 10 November) and Anthropic published “The components of agent memory.” I discuss both in a companion post. On 13 November 2026, Google announced that U.S. Gemini app users can connect Google Wallet for read-only, spending-informed answers, and that Memory Summary lets users view, edit or remove what Gemini remembers: https://blog.google/innovation-and-ai/products/gemini-app/google-wallet-gemini-app/.
How I am reading the week If you only have time for a few actions: upgrade mcp-server-git to 2026.11.11 or later, Inspector to 2.15.0, and AWS Pricing MCP Server to 1.0.31 (and ContextForge to 1.1.1 if you run it). Then read the AAIF post as a map of where skills, plugins and MCP are heading, and treat the Ignite Agent 365 agenda as a signal that tool-level governance is becoming an enterprise control plane.
My own tools sit on the same problems this week's news keeps surfacing. mcp-pordego scans MCP setups; agentscrub redacts secrets and PII from agent transcripts; Ambient Memory is a local-first context layer; Agent Replay Lab records and mock-replays tool I/O. Project pages are at https://jarvisosorio.com/open-source/ Companion posts: agent security at /blog/jarvis-nelson-osorio-mcp-agent-security-week-november-2026/ and memory and observability at /blog/jarvis-nelson-osorio-agent-memory-observability-week-november-2026/.