Tech Radar - 2026-10-16 18:00

Priority Today

Together AI is serving Muse Spark Open free until October 18

Together AI is offering Meta's Muse Spark Open at no cost through October 18 at 12 p.m. PT, as a promotional model id on its serverless inference API. The promo row bills input and output at $0.00 for the window, against a standard Muse Spark Open rate of $1.25 per million input tokens and $4.25 per million output tokens on the same page.

Why it matters: The 816B open-weight model that Meta released on October 15 can be evaluated on your own tasks without paying for the two days the offer lasts. What changed: A promo row marked "valid until 10/18/2026" prices input and output at $0.00, alongside the normal paid Muse Spark Open id. For me: Point an OpenAI-compatible client at Together's endpoint with the promo model before the deadline. The offer ends rather than converting into a standing free tier.

Tags: #topic/ai-economics #topic/ai-models #entity/meta #entity/together-ai #technology/llm-api #event/free-tier #relevance/free-access #learning/hands-on

Sources:

GitLab's MCP server is generally available with GitLab 19.5

GitLab made its MCP server generally available on October 15 with GitLab 19.5, exposing code, merge requests, work items and pipelines to any MCP-compatible agent. Each action runs under the user's existing GitLab permissions, and an agent can also start and follow GitLab Duo Agent Platform flows.

Why it matters: An agent reaches the whole software lifecycle through one MCP server with focused tools instead of guessing at REST endpoints and burning context on extra calls. What changed: New tools include add_branch, get_repository_file and semantic_search for code; get_pipeline, list_pipelines and get_artifact_file for pipelines; list_groups and opt-in list_wiki_pages for orientation. get_merge_request now returns commits, diffs, pipelines, conflicts or notes, toolsets select which groups are exposed, and Duo handoff tools (start_duo_session, get_duo_session, send_duo_session_input) are on by default. It is available on Free, Premium and Ultimate, on GitLab.com, self-managed and dedicated, with per-user rate limits of 60 calls per minute on Free and 600 on Premium and Ultimate. For me: Connect it from Claude Code, Cursor, GitHub Copilot, Codex, Gemini, Kiro or Zed using the documented setup; the Free-tier limit of 60 calls per minute is the number to watch.

Tags: #topic/ai-coding #topic/ai-agents #entity/gitlab #technology/mcp #event/tool-release #relevance/developer #relevance/ai-workflow #learning/hands-on

Sources:

IntentKit's CVE-2026-76066 let unauthenticated callers run owner-only tools

VulnCheck published CVE-2026-76066 on October 15, an authorization bypass (CWE-862) in IntentKit, Crestal Network's open-source agent framework, rated CVSS 8.7. IntentKit narrows the tools it advertises to the model per request but builds the executor from the full private registry, so a call naming a de-advertised tool still runs, including a wallet transfer.

Why it matters: It shows that hiding a tool from the model is not an authorization boundary; the check has to sit where the tool is executed, or a de-advertised tool stays reachable. What changed: IntentKit 2.6.3 and every earlier release are affected, no fixed version is named, and the latest release is v2.6.3 from June 14, 2026. The bypass is reachable through indirect prompt injection or by replaying checkpoint and history on a reused thread, and no tool carries an approval gate. The always-present cdp_native_transfer wallet tool plus DeFi swap, posting and memory-write tools are in reach. For me: If you expose an IntentKit agent to public users or feed it untrusted content, build the executor's tool set from the same context-filtered list shown to the model, and gate state-changing tools until a patch lands.

Tags: #topic/security #topic/ai-agents #entity/intentkit #technology/agent-security #event/security-issue #relevance/developer #relevance/ai-workflow #learning/understand

Sources:

Vercel launched Token Pass, $30 a month for $80 of GLM 5.3 and Kimi K3 credit

Vercel opened early access to Token Pass on October 15: for $30 per team each month, a Pro team gets $80 of credit for the featured open-weight models GLM 5.3 (zai/glm-5.3) and Kimi K3 (moonshotai/kimi-k3) through AI Gateway, served by the Token Pass provider (currently Wafer) with zero data retention.

Why it matters: It is a fixed-cost way to run two frontier open-weight models inside any AI Gateway-connected coding agent, with a hard monthly spend cap and no automatic overage. What changed: Credit resets to $80 each billing period and does not roll over; when it runs out, Token Pass requests fail with a 403 until the next period starts. A Token Pass API key works only with the featured models, and fast variants are not covered. For me: Pro teams subscribe from the AI Gateway Token Pass tab and point an agent at GLM 5.3 or Kimi K3; keep a regular AI Gateway key for every other model.

Tags: #topic/ai-economics #topic/ai-coding #entity/vercel #technology/llm-api #event/pricing-change #relevance/cheap #relevance/ai-workflow #learning/hands-on

Sources:

Free and Cheap Access

NVIDIA's Nemotron 3.5 Lightning is free on Vercel AI Gateway

NVIDIA's Nemotron 3.5 Lightning is available for free on Vercel AI Gateway as nvidia/nemotron-3.5-lightning-free, served by NVIDIA, per Vercel's October 14 changelog. It is an open-weight reasoning and chat model aimed at long-running agents, sub-agent work, chatbots and RAG.

Why it matters: A 1M-context, tool-calling open model at zero cost lowers the price of wiring a capable model into an agent you are still evaluating. What changed: The model has 30B parameters with 3B active per token, a hybrid Mixture-of-Experts design interleaving Mamba-2 and MoE layers with a few attention layers, a 1M-token context window, and support for tool calling and structured outputs. Vercel notes the provider may retain prompts for model improvement. For me: Set the model to nvidia/nemotron-3.5-lightning-free in any AI Gateway-connected agent. Treat it as not zero-retention, unlike the Token Pass models.

Tags: #topic/ai-economics #topic/ai-models #entity/nvidia #entity/vercel #technology/context-window #event/free-tier #relevance/free-access #learning/hands-on

Sources:

Vercel changelog: Nemotron 3.5 Lightning now available on AI Gateway for free

AI Coding and Agents

Copilot CLI 1.0.98 adds Gemini 4 Argon and agent-scoped declarative hooks

GitHub Copilot CLI shipped 1.0.98 on October 14 and 1.0.97 on October 13. 1.0.98 adds Gemini 4 Argon support, while 1.0.97 adds agent-scoped declarative hooks for custom agents and makes managed permission rules read-only in /config unless set by your organization.

Why it matters: The permission and hook changes sit on the trust boundary of an unattended coding agent, and the marketplace retry path now uses your GitHub token. What changed: Autopilot now denies a preToolUse hook's "ask" decision instead of stopping on a permission dialog nobody is watching. Remote plugin marketplace and plugin fetches retry once with your signed-in GitHub token after an HTTPS auth failure. /usage shows billed AI Credits for token-billed sessions, and saved settings survive one invalid config.json entry. For me: Update to 1.0.98. If you run autopilot, a hook that returns "ask" now denies rather than prompting, so review any hooks that relied on the old behavior.

Tags: #topic/ai-coding #topic/developer-tools #entity/copilot #entity/github #technology/agent-security #event/tool-release #relevance/developer #learning/hands-on

Sources:

GitHub Copilot CLI changelog

Agent Capability

Microsoft added always-on triggers to the GitHub Copilot harness in Copilot Studio

Microsoft announced on October 15 that triggers are available in the GitHub Copilot harness in Copilot Studio, letting an agent built on the harness start a cycle from an external signal such as an email, or from a set cadence, with no prompt. The same announcement added agent accounts and business-process memory so an agent persists beyond one person's role.

Why it matters: An agent that wakes on an external event rather than a user turn is the difference between a chat tool and an always-on process worker. What changed: Triggers, agent accounts and business-process memory extend Copilot Studio toward organization-owned "digital teammates" that run in the apps and services a team already uses. The capabilities are rolling out through the Frontier program. For me: If you build agents with the GitHub Copilot harness, triggers remove the need for a human to start each run. Access is gated by the Frontier program for now.

Tags: #topic/ai-agents #topic/ai-coding #entity/microsoft #entity/copilot #technology/agent-orchestration #event/tool-release #relevance/ai-workflow #learning/understand

Sources:

Open Source Momentum

Anthropic's claude-plugins-official is the sanctioned Claude Code plugin directory

Anthropic maintains anthropics/claude-plugins-official, described as its official, Anthropic-managed directory of high-quality Claude Code plugins, split into internal Anthropic plugins and third-party plugins from partners and the community. The repository is Apache-2.0 and shows roughly 39.8k stars and 4.4k forks.

Why it matters: Plugin distribution is where an agent's capabilities and its supply-chain risk both come from, and an official directory becomes the default place users install from. What changed: Plugins install with /plugin install {name}@claude-plugins-official or from /plugin > Discover. Plugin names are immutable slugs with a renames map for migration, and skill-bundle plugins are supported. Anthropic warns it does not control plugin contents and cannot verify they work as intended. For me: Install from this directory when you want an Anthropic-managed catalog, but treat every plugin as untrusted, since Anthropic explicitly disclaims verification of what a plugin includes.

Tags: #topic/ai-coding #topic/open-source #entity/anthropic #entity/claude-code #technology/agent-skills #event/github-traction #relevance/developer #learning/awareness

Sources:

GitHub: anthropics/claude-plugins-official

CLI-Anything turns desktop applications into agent-native CLIs

HKUDS/CLI-Anything generates agent-native CLI harnesses for software that has no API, such as Blender, QGIS, GIMP, FreeCAD, LibreOffice, Sketch and Zoom. The project is Apache-2.0 and shows roughly 53.2k stars and 4.8k forks.

Why it matters: It is a different integration path from MCP: instead of building a server per tool, it wraps existing software in a CLI with structured output an agent can call. What changed: One command installs the hub (pip install cli-anything-hub), then cli-hub install <name> adds a harness that returns JSON. A meta-skill lets an agent discover and install the right CLI itself, each harness ships a SKILL.md, and the registry adds new apps continuously. For me: Run pip install cli-anything-hub to browse and install, or add the meta-skill so your agent picks a harness on its own. Some CLIs wrap real desktop software that must also be installed.

Tags: #topic/open-source #topic/ai-agents #entity/cli-anything #technology/agent-skills #event/github-traction #relevance/developer #relevance/discovery #learning/understand

Sources:

New Concepts

Tool authorization must match the advertised tool set

When an agent framework narrows the tools it shows the model per request but validates the executor against the full registry, the narrowing is cosmetic: a tool the model was never offered can still be invoked by name. IntentKit's CVE-2026-76066 is the clearest case this week, and the same advertised-versus-executed gap underlies a run of recent approval-bypass advisories in coding agents, where a permission check reads one tool list while the runtime runs another.

Why it matters: It is a recurring class of agent vulnerability. A filter applied only to the model's view buys no security; the authorization decision has to be enforced where the tool actually runs. What changed: The pattern showed up in IntentKit this week and in earlier advisories against several coding agents, so it is worth recognizing as a design rule rather than a one-off bug. For me: When you build tool dispatch, derive the executor's tool set from the same context-filtered list the model saw, and reject any call whose name is not in that turn's advertised set. Add a human-in-the-loop gate on state-changing tools for public or shared agents.

Tags: #topic/security #topic/ai-agents #technology/agent-security #event/security-issue #relevance/ai-workflow #learning/understand

Sources:

Tag Index

topic

#topic/ai-agents #topic/ai-coding #topic/ai-economics #topic/ai-models #topic/developer-tools #topic/open-source #topic/security

entity

#entity/anthropic #entity/claude-code #entity/cli-anything #entity/copilot #entity/github #entity/gitlab #entity/intentkit #entity/meta #entity/microsoft #entity/nvidia #entity/together-ai #entity/vercel

technology

#technology/agent-orchestration #technology/agent-security #technology/agent-skills #technology/context-window #technology/llm-api #technology/mcp

event

#event/free-tier #event/github-traction #event/pricing-change #event/security-issue #event/tool-release

relevance

#relevance/ai-workflow #relevance/cheap #relevance/developer #relevance/discovery #relevance/free-access

learning

#learning/awareness #learning/hands-on #learning/understand