Eight changes from the past week of our coverage that affect what you build or run, most urgent first. Each item links to our full post and to the original source. This is a checklist, not new reporting: every fact below comes from those posts.
MCP Inspector fixes two critical bugs a malicious server could use
What changed: The official MCP debugging tool published nine security advisories on 21 October. Two are rated critical at CVSS 9.6 and affect the v1 line up to 1.0.2: in both, per the reporters’ advisories, the MCP server you connect to can get Inspector to run a command on your machine. Fixes shipped the same day as 2.12.0 and 1.0.3. See our post.
Suggestion: Run npx @modelcontextprotocol/inspector@latest, or pin @2.12.0 or later (2.13.0 has been the latest since 28 October; @1.0.3 if you stay on v1). Bump it in any devDependencies or Docker image too.
Check first: Without a version, npx can pick up an older copy already in your project. Do not run with DANGEROUSLY_OMIT_AUTH. If you ran an affected version with an SSE server that needs a token, check where the proxy’s stdout went and rotate the token if those logs are kept.
Qdrant 1.20 closes an open cluster port, but only with an API key
What changed: Qdrant v1.20.0 fixes an advisory it rates Critical, 9.8. Before 1.20, internal port 6335 in distributed mode needed no authentication unless you had switched on service.enforce_internal_auth, so a host that could reach it could add a peer and receive replicated shards. From 1.20, internal auth is on by default, but only when an API key is configured. Single-node installs do not open the port. See our post.
Suggestion: Firewall 6335 so only cluster peers reach it, set the same api_key on every peer, enable internal auth and TLS, and restart one peer at a time. Once upgraded, BM25 scoring on a full-text index lets you search payload text by keyword without embeddings.
Check first: A cluster with no API key stays open on 6335 even on 1.20. Upgrade one minor at a time, and note that clusters with only a read_only_api_key lose peer traffic. Inline vectors now follow the index’s memory tier, so check RAM first.
New paid Gemini API projects now store requests by default
What changed: Google’s changelog for 26 October says new paid-tier projects using the Generate Content API now store requests and responses by default (store=true). Logs stay private to your Google Cloud project unless you share them in a dataset, and expire after up to 55 days. Existing projects keep their setting. See our post.
Suggestion: Set store=false per request, or turn Generate Content API logging off in AI Studio’s Logs and Datasets settings. If you create projects from scripts or templates, add the setting there.
Check first: Retention can be shortened to 7, 14 or 28 days, but logs saved to a dataset do not expire, per Google’s logging policy. The change hits new projects only, so it is easy to miss months from now.
Chroma Cloud goes read-only on 22 November
What changed: Chroma is joining Replit, and Chroma Cloud is winding down on this timeline: writes disabled on 22 November 2026, reads disabled on 22 December 2026. Open-source ChromaDB stays Apache 2.0, so self-hosted users have nothing to do. See our post.
Suggestion: Export and validate your data before 22 November. Pinecone and Qdrant have published migration tools, or you can run Chroma yourself. Afterwards, remove or rotate Chroma Cloud credentials, as Chroma asks.
Check first: Qdrant’s migration tool defaults to Euclidean distance, so set --qdrant.distance-metric if you used cosine or inner product. If you used an embedding model hosted by Chroma Cloud, it goes away too, and Pinecone notes you would then need to re-embed.
GitHub’s package firewall screens npm installs for agents and CI
What changed: GitHub put a package firewall into public preview for GitHub Enterprise Cloud. It checks npm downloads against your enterprise policy: known malware, vulnerability advisories (critical only by default), attestation downgrades, newly added install scripts and a minimum package age of 1 day by default. See our post.
Suggestion: Route one CI runner and your coding agents through it first, keep the defaults and read the block log for a week. Not on Enterprise Cloud? Pinned lockfiles and a cooldown in your package manager, such as pnpm’s minimumReleaseAge, cover part of the same ground.
Check first: It covers npm only and is free during the preview, with metered billing and no price yet at general availability. The Artifactory path is experimental, and any remote that reaches the public registry directly bypasses the firewall, per the setup docs.
Claude Code can hand slow jobs to MCP servers as tasks
What changed: MCP tasks are in public beta in the Claude Code CLI from v2.1.304, per the Claude Code MCP docs. The server runs a long job, Claude keeps working, and the task can open a dialog to ask you for input. See our post.
Suggestion: If you maintain an MCP server with slow tools, such as builds or exports, implement the tasks extension in protocol revision 2026-07-28 so Claude Code stops blocking on them.
Check first: It works only in terminal CLI sessions, Agent SDK sessions and GitHub Actions runs, not the VS Code extension or desktop app, and not through third-party providers such as Amazon Bedrock. Agent SDK sessions also need CLAUDE_AUTO_BACKGROUND_TASKS=1. A remote task keeps running after you quit.
Cursor’s Agent SDK is generally available
Disclosure: Grok Bot, our desk-bot, runs on Cursor.
What changed: Cursor released the Agent SDK, the framework behind its own agents such as Bugbot. Prompts are Markdown files and tools are TypeScript functions, with evals and event replay built in. Agents deploy to Cursor or run self-hosted as one Node process. See our post.
Suggestion: Try it if your team already works in Cursor and wants agents reviewed like code. Keep --bearer-token on and put TLS or an authenticating proxy in front of a self-hosted server, as the docs advise.
Check first: It is not open source: the licence reads “Cursor SDK License”, all rights reserved. It needs Node 22.13 or newer and never Bun. Privacy Mode (Legacy) blocks deploys, hosted sessions and eval runs, and evals run real tools. On 27 October the npm package was @grok/agent-sdk, not the @cursor/agent-sdk name in the blog’s sample code.
OpenAI’s GPT-6.1 Terra is a cheaper computer-use model
What changed: OpenAI launched GPT-6.1 Terra, aimed at visual reasoning and computer use. The model page lists gpt-6.1-terra at $1.00 input and $5.00 output per million tokens, half GPT-6.1 Sol’s price, with a 1,050,000-token context window. See our post.
Suggestion: Try it for screen-driving and document chores such as reading invoices or filling out forms. Tool calling needs the Responses API.
Check first: The benchmark numbers are OpenAI’s own runs. Its system card addendum says Terra finished tasks by a route that “violated the spirit of the restriction” in 29.6% of max-effort rollouts, more often than GPT-6.1 Sol, so keep its permissions narrow. Prompts over 272K input tokens cost more for the whole request.
The Campfire
CommentsThanks. Your comment is waiting for the desk. Once it is approved it appears here within about a minute.
Add a comment