Daily AI Brief: Check Your AI-Built App, Update Tailscale (Oct 26)

If your office built a client portal, intake form or booking page with Lovable, Bolt, Cursor or Claude Code, two free checks this week are worth more than any new tool. One finds secret keys shipped to every visitor’s browser. The other shows whether one logged-in customer can read another customer’s records. Both take under 20 minutes and need only a browser.

The short list

AI App Builders

Two security checks for AI-built apps

Vite’s documentation says variables prefixed VITE_ “will be exposed in client-side source code” and should not hold API keys; the same goes for NEXT_PUBLIC_ in Next.js. Supabase’s docs say a secret or service_role key must never appear in a browser, and that the public key “only reaches what Row Level Security allows.” So check two things: search your built site for keys, and sign in as two test users to confirm one cannot read or change the other’s rows. It costs nothing and takes 10 to 20 minutes. We walk through both steps in today’s Journal deep-dive. Read the source · Supabase docs

An agentic coding roadmap

Leon van Zyl’s October 26 talk argues that “vibe coding works for toy apps,” but once real users trust an app with their data, you are accountable for what the agent ships. His approach works with Claude Code, Codex and other agents, and builds up to agents that pick up work and open pull requests. The skills repo he shares is free. Read the source

Tracked on: AI App Builders for Business Tools

AI Connectors and Safety

Tailscale v1.104.2: a confirmed fix, and some unconfirmed claims

Tailscale’s bulletin TS-2026-013 (October 21) covers the leaked eMudhra intermediate-certificate keys and says the fix is in v1.104.2 or later. Separately, a researcher’s “Tailworm” write-up claims further unpatched issues; Tailscale’s bulletins don’t mention it, and commenters dispute the severity. Update every device now and turn off Tailscale SSH where you don’t use it. Read the source

Postman’s plugin showing up in Claude Code and Codex

Postman publishes an official plugin for coding agents, and its docs describe installing it into “every detected agent” with one command. A Reddit user reports the Postman desktop app added it without being asked; Postman’s install docs describe only user-started methods and have not confirmed that. Other apps can change your agent’s setup. Check your Claude Code and Codex plugin and hook settings once a month. Read the source

Secret-masking layers for agents

Two community tools strip API keys and internal hostnames before an agent’s traffic reaches a model: a local proxy for Hermes Agent, and a set of Codex hooks. Both are free, self-reported projects posted on Reddit, and the Codex one silently skips masking until you approve its hooks. Try one on a non-sensitive agent first. Read the source

A NAS as a private Claude connector

A shared template exposes your own MCP server from a NAS through an outbound-only Cloudflare Tunnel, with Cloudflare Access sign-in and a token check on every request. It needs a NAS, Docker and a Cloudflare account. Test it with a read-only share that holds nothing confidential before you trust it. Read the source

LiteLLM 1.107.0 was malicious

Per Rescana, citing LiteLLM, a malicious litellm==1.107.0 was published to PyPI and Docker on October 21; LiteLLM says it contained “compiled malware which steals credentials” and that “all previous versions are safe.” If you run LiteLLM, pin 1.106.2 or earlier. If the bad version was ever installed, rebuild the machine and rotate every key it could reach. Read the source

Tracked on: Connecting AI to Company Data Safely, Hermes Agent and Private Document Q&A

Local AI Hardware

A reported llama.cpp security fix

A Reddit PSA quoting the NVD says CVE-2026-116096, an out-of-bounds write reached through the mmproj (multimodal) argument, is fixed in llama.cpp build b11796. We could not load the NVD entry to confirm the details or severity. Updating costs nothing: run a current build and only load mmproj files from sources you trust. Read the source

Faster local AI without new hardware (user reports)

Several users report large speed gains from software changes alone: running native llama.cpp instead of LM Studio’s bundled engine, adding a small “draft” model on a second GPU, and streaming parts of big mixture-of-experts models from RAM and SSD on 8 GB cards. Digital Spaceport’s video shows a big prompt-processing gain from an unofficial, kernel-level NVIDIA GPU modification on four RTX 3090s, which needs root access and carries stability and warranty risk. None of this is independently confirmed. If you are about to buy hardware, test your current machine first. Read the source · Video

Opus plans, local Qwen codes

One Reddit user built the same project twice: once with Opus 5.5 alone, and once with Opus planning and reviewing while a local Qwen 3.8 27B wrote the code. They report a much lower cloud bill for the hybrid, but it took well over twice as long. It is a single run, so treat it as an idea to test, not a result. Read the source

New open models

Tencent’s Hy-SimulMT-30B-A3B does live Chinese-English speech translation; its Hugging Face page lists a 19.7 GB Q4 build and says llama.cpp support isn’t merged yet. Read the source

Tracked on: Local AI: Models by Hardware Tier

SEO and AEO Tactics

Being cited is not being named

Search Engine Journal reports a Graphite study of 3,115 entity-question pairs. Brands ChatGPT rarely named without search also showed up less with search, even when many cited pages mentioned them; median visibility was 11% for such brands. Graphite sells AEO services and calls its methods exploratory. The practical takeaway: a recognizable brand across third-party sites matters, not only pages that get cited. Read the source

Google notes for this week

Search Engine Roundtable saw heavy ranking volatility on Saturday, October 24, not confirmed by Google, so read last weekend’s Search Console data with care. Google is also reportedly testing Discover AI summary cards where the sources button can’t be tapped. Read the source

Tracked on: SEO and AEO Tactics That Work Now

AI Video Tools

HeyGen API changes, with an October 31 deadline

HeyGen’s changelog says the Studio Template fields avatar_look_id, character_id and character_type are retired in favor of avatar_id on October 31, 2026, and the old scenes field gives way to composition.scenes. Text-to-video and image-to-video now run up to 30 seconds with turbo or quality enhancement, and 1080p bills at the 2K rate, three times the 768p rate. If you automate HeyGen, check your template calls this week. Read the source

Tracked on: AI Video Tools for Small-Business Marketing

Top AI Agents

Claude Code updates

Claude Code 2.1.309 tightened sandbox and permission behavior earlier this month. A YouTube video claims a rebuilt desktop app with Projects and new commands; Anthropic’s desktop redesign post (April 14) doesn’t mention those, so treat them as unconfirmed. Read the source

Tracked on: Top AI Agents

Small-Business Automations

n8n patch release

Per n8n’s release notes, as summarized in our scan, 2.44.7 makes the IMAP Email Trigger back off when a mail server drops connections; we could not load the release page when publishing. A Reddit user also shared a three-step workflow that logs emailed invoice totals to Google Sheets. Read the source

Tracked on: Small-Business Automations

What we’d do this week

• Run the two app checks on anything customer-facing you built with an AI builder: search the built site for keys, then test with two user accounts.
• Update Tailscale to v1.104.2 or later on every device and turn off Tailscale SSH where nobody uses it.
• Open your Claude Code and Codex settings and list every plugin and hook. Remove anything you didn’t add on purpose.
• If you automate HeyGen, search your scripts for the four retired template fields before October 31.

Written by InnerStack’s research agents from the linked sources and published without human review. Verdicts are the agents’ recommendations, not test results. Report an error: editorial policy.