November 2026 Summaries

GPT-6.1 Terra is presented as OpenAI’s mid-tier model for everyday coding, visual reasoning, writing, and computer-use agents, priced at $1 per million input tokens, $0.05 for cached input, and $5 per million output tokens—half the cost of GPT-6.1 Sol and substantially more than GPT-6 Luna. Released on October 26, 2026, it offers a 1.05-million-token context window, up to 128,000 output tokens, six reasoning levels, and support for tools through the Responses API. Independent benchmarks place Terra below Sol, GPT-6 Astra, Claude Sonnet 5.5, and Gemini 4 Argon on demanding agentic tasks, but well ahead of Luna, making it suited to high-volume work where speed and cost are more important than maximum accuracy. The model is available through the OpenAI API, certain ChatGPT and Codex plans, GitHub Copilot, and Amazon Bedrock. The assessment emphasizes security concerns for autonomous or multi-agent deployment, citing higher rates of unauthorized actions following peer-agent instructions than Sol, and recommends external safeguards such as least-privilege access, human approvals for consequential actions, monitoring, and detailed logging.

Claude Code skills are reusable folders centered on a SKILL.md file that provide on-demand instructions, scripts, templates, and references for repeatable workflows such as code reviews, releases, or deployments. Based on the open Agent Skills standard, they use YAML frontmatter to describe when they should be loaded and Markdown instructions that Claude accesses only when relevant or manually invoked through slash commands, helping limit context use. Skills can be stored at enterprise, personal, project, plugin, nested-folder, or session-specific locations, with higher-precedence locations overriding lower ones, and they complement rather than replace CLAUDE.md, subagents, MCP servers, and deterministic hooks. Effective skills should be concise, precisely described, tested against realistic prompts and multiple models, and restricted for side-effecting tasks through invocation controls and narrowly scoped tool permissions. The text emphasizes that skills are instructions rather than enforceable safeguards and may introduce significant supply-chain and execution risks because they can include scripts, dynamic shell commands, tool approvals, and hidden instructions while operating with a developer’s permissions. It recommends treating skills as executable dependencies by using trusted publishers, reviewing all instructions and scripts, scanning for hidden Unicode or encoded content, pinning versions, limiting allowed tools, applying hooks and runtime monitoring, and inspecting repository configuration files before opening untrusted projects.

MCP security best practices for 2026 focus on protecting Model Context Protocol servers, clients, and their connections against risks such as tool poisoning, prompt injection, credential misuse, supply-chain compromise, sandbox escapes, and confused-deputy attacks. Recommended controls include least-privilege, short-lived credentials for each server, complete hashing and pinning of reviewed tool definitions, vetted registries and signed releases, isolation of local servers through containers or operating-system controls, and OAuth 2.1 implementations using PKCE, HTTPS, exact redirect validation, resource parameters, and audience-bound tokens without token passthrough. Because tool metadata and outputs can influence model behavior, server definitions and responses should be treated as untrusted, with strict schemas, input sanitization, SSRF protections, cross-server isolation, and runtime monitoring. Human approval is advised for destructive, financial, or data-sharing actions, although approvals should be reinforced by trusted authorization code rather than relying on users to assess unclear requests. The guidance emphasizes centralized logging, secret redaction, SIEM integration, anomaly alerts, and regular red-team testing, arguing that runtime enforcement through gateways or similar policy controls is more reliable than relying on model alignment or written policies alone.