AI Frankly

A note before you listen: this episode's deep dive was produced in NotebookLM. NotebookLM is our executive producer. It also mispronounces Ollama as "Alama" about forty times. We use AI to talk about AI. When the AI hits its limit, we point at it instead of hiding it.

---

Three recipes for a voice assistant that doesn't have to phone home.

Recipe 1 takes ten minutes — a browser-based Jarvis you can run today.

Recipe 2 takes thirty minutes — a one-command local install with Ollama.

Recipe 3 takes a weekend — Home Assistant Voice PE, replacing Alexa entirely.

I built one. I read the receipts on two more. Here's what each actually costs in time, money, and trade-offs.

Read the full article: https://aifrankly.substack.com/p/the-local-jarvis

CHAPTERS:

00:00 - The Smart Home Illusion

03:30 - Recipe 1: The Browser Jarvis

07:45 - The Browser API Privacy Trap

12:00 - Recipe 2: The One-Command Jarvis

15:30 - Why It Knows When You Stop Talking

18:15 - The Hardware Reality Check

22:00 - Recipe 3: The Whole-Home Jarvis

26:45 - The Frustrating Five-Second Delay

30:00 - The Honest Bill

32:15 - The Partner Test

34:30 - Local Is Not a Security Policy

36:45 - Auditing Microphone Vulnerabilities

40:30 - High-End and Cloud Alternatives

44:15 - Am I Tony Stark?

RECEIPTS:

- Ollama: https://ollama.com

- Home Assistant Voice PE: https://www.home-assistant.io/voice_control/

- Kokoro TTS: https://github.com/hexgrad/kokoro

- NetworkChuck's local AI tutorials: youtube.com/@NetworkChuck

- Eddie Chen's voice assistant builds: youtube.com/@EddieChen

- Smart Home Solver's HA Voice PE testing: youtube.com/@SmartHomeSolver

AI Frankly. Built from parts. Owned outright.

MCP 202: The Missing Control Layer Between Your Agents and Your Tools.

Security researchers have a name for what's happening inside most enterprise AI deployments right now: NeighborJack.

This episode breaks down what NeighborJack is, why vendors are racing to sell you MCP Security Gateways, and the four questions every operator should ask about any MCP server they're running.

Full article: https://aifrankly.substack.com/p/someone-needs-to-sit-between-your-b31

Watch on YouTube: https://youtu.be/TKiDj4vH4Vc

Subscribe free at aifrankly.substack.com.

AI Frankly: Build the layer or buy the breach.

Half a million lines of source code leaked. A federal court fight the same week. And the thing the internet actually remembered was an ASCII owl named Zornix.

This episode covers what EP.15 and EP.16 didn't -- the cultural artifact that came out of the worst press week in Anthropic's history.

Download the Python script (see all 18 buddies, free): https://tinyurl.com/zornix

Read the full article: https://aifrankly.substack.com/p/the-week-anthropic-leaked-the-internet

EP.15 -- Anthropic Did It Again: https://aifrankly.substack.com/p/anthropic-did-it-again

EP.16 -- The Blueprints Are Public Now: https://aifrankly.substack.com/p/the-blueprints-are-public-now

00:00 - Introduction: The Dissonance 03:00 - Timeline of the Leak 07:00 - The Mechanics of Minification 12:00 - The Discovery (30 Seconds) 16:00 - Unpacking KAIROS: The Autonomous Daemon 21:00 - Undercover Mode 26:00 - The Compaction Attack Vector 30:00 - The Bash Security Parser Differential 34:00 - The Axios NPM Supply Chain Collision 38:00 - Mitigation Protocols 42:00 - The Pentagon Lawsuit and Responsible Scaling 45:00 - The Verification Agent 48:00 - The Phenomenon of Zornix 50:00 - Conclusion

Subscribe free on Substack: https://aifrankly.substack.com Apple Podcasts: https://podcasts.apple.com/us/podcast/ai-frankly/id1873177211 Spotify: https://open.spotify.com/show/6iEuQyxwLeUDsR67QleSTn

AI Frankly:

aifrankly.com

Anthropic refused defense contract terms that would have allowed military use of Claude. A federal judge called the Pentagon's response classic First Amendment retaliation. Seven days later, the complete source architecture of Claude Code was permanently mirrored across GitHub.

This is the full deep dive. 41 minutes covering the Pentagon lawsuit, Judge Rita Lin's ruling, KAIROS, Undercover Mode, the compaction attack vector, and what it means for enterprise IT practitioners right now.

What we cover:

- The Pentagon dispute and Judge Rita Lin's ruling

- KAIROS -- the fully built autonomous daemon nobody announced

- Undercover Mode and the irony at the center of the story

- The compaction attack vector and context poisoning

- The bash security parser differential

- The Clean Room Clone and the IP legal paradox

- The IPO stakes and what this means for enterprise trust

Watch on YouTube: https://youtu.be/0L2isrk88y0

Read the full article: https://open.substack.com/pub/aifrankly/p/the-blueprints-are-public-now

Find everything at aifrankly.com

On March 31, 2026, an intern at a blockchain security firm checked the npm registry for the newest release of Claude Code. Within 30 seconds he found a 59.8MB source map file that should never have shipped to production. Inside: 512,000 lines of TypeScript, 1,900 files, and the complete architecture of one of the most important AI developer tools ever built.

This is the full deep dive. 45 minutes covering the mechanics of how it happened, what the leaked source code actually revealed, and the concurrent Axios supply chain attack that hit the same morning.

What we cover:

How a single missing line in a config file exposed the entire codebase

KAIROS: the fully built autonomous daemon mode Anthropic never announced

The Buddy virtual pet system and the hex-encoded duck

Undercover Mode and the grand irony at the center of this story

The Axios npm supply chain attack and who is actually at operational risk

The operator verdict: Watch, Act Now, Adopt

Watch on YouTube: https://youtu.be/XGv8sW2NS0k

Read the full article: https://open.substack.com/pub/aifrankly/p/anthropic-did-it-again

Find everything at aifrankly.com

Anthropic shipped six major products in March 2026. A federal lawsuit over autonomous weapons. And a quietly revised safety pledge with no hard stop. All in the same 30 days.

In this episode we unpack the full operator picture -- what shipped, what it means for enterprise security, and what every IT practitioner needs to know before deploying any of it.

In this episode:

Claude Computer Use and the endpoint security risk nobody has filed a ticket for yet

Claude Code Auto Mode and the shadow IT threat of Channels

The death of RAG pipelines with 1M context at standard pricing

The $100M partner network with the big four consulting firms

The RSP revision that removed the hard stop

The Pentagon lawsuit and what it means for your vendor risk assessment

Operator Verdict: Adopt, Watch, Skip -- exactly what to do with each update

Find everything at aifrankly.com Watch on YouTube: https://youtu.be/19wv2Pqa0Eg Listen on Apple Podcasts and Spotify -- search AI Frankly

Nobody hacked Anthropic. A misconfigured CMS left roughly 3,000 unpublished assets publicly accessible. One described a new model nobody was supposed to see yet. Cybersecurity stocks dropped billions in a single session. Not a breach. A config error.

In this episode we unpack the full anatomy of the incident, what actually leaked, why Wall Street panicked, and what it means for every enterprise IT team managing content and documentation systems right now.

In this episode:

The anatomy of the CMS misconfiguration

What leaked: Claude Mythos and the Capybara tier

Why cybersecurity stocks dropped billions in one session

What Anthropic actually shipped in March 2026

Claude Computer Use reaches general availability

The 15-minute rogue audit every IT team should run today

Find everything at aifrankly.com Subscribe free on Substack Watch on YouTube: https://youtu.be/4zheW7QJ4k4 Listen on Apple Podcasts and Spotify -- search AI Frankly

MCP hit 30 CVEs in 60 days. 38 percent of public MCP servers

had zero authentication. The protocol works. The governance

layer doesn't exist yet.

This episode covers the real threat model for MCP in

production: three villains, five layers of defense, and a

six-question self-audit you can run against any MCP server

today.

MCP didn't break your security model. It exposed that you

didn't have one.

Topics covered:

- The incident that should have been an incident report

- The Content Injector, the Supply Chain Attacker, and the Over-Helpful Agent

- Five layers of defense: identity, supply chain, isolation, policy, and monitoring

- The governance maturity model: Stage 0 through Stage 3

- The six-question self-audit

Read the full article: https://aifrankly.substack.com/p/mcp-201-the-governance-deficit

Full MCP Security Series: https://aifrankly.com/mcp-security

AI Frankly: Are We Having Fun Yet!

Bonus Episode - Season 1

Anthropic didn't have a week. They had a product conference that nobody announced.

This deep dive breaks down everything that shipped in March 2026, sorted by who it actually affects: everyday Claude users, developers building with Claude Code, and enterprise operators responsible for securing all of it.

What we cover:

Persistent memory and the governance nightmare

Sonnet 4.6 and the 1M token context window

Why RAG pipelines are becoming obsolete

Claude in Chrome and the Cowork native host conflict

The /loop command and runaway agent scenarios

Claude Code Channels and the Telegram attack vector

Voice mode and generating bad code faster

Claude Code Security grading its own homework

Cowork on Desktop and why the VM sandbox fails

Superpowers, Dispatch, and awesome-agent-skills

The IT velocity gap that is now permanent

The memory test: ask Claude what it remembers

Free resource mentioned in this episode: claudehq.app - no account required

Full article: https://open.substack.com/pub/aifrankly/p/this-already-hit-production-your

AI Frankly: The memo your IT team didn't get.

Here are the show notes:

MCP 103: I Gave Claude Code Someone Else's Tools (Full Deep Dive) Season 1, Episode 11

This week I wired three public MCP servers into Claude Code and gave it one sentence. It touched GitHub, n8n, and Notion without being asked which ones to use. No ticket filed. No audit trail.

This episode covers the full lab report: what I connected, what Claude did without being told, and why the governance conversation is happening after the fact in most enterprises right now.

Topics covered:

GitHub MCP: read-only access and unprompted anomaly detection

n8n MCP: autonomous workflow chaining across three automations

Notion MCP: persistent memory that survives the conversation window

What happened when all three were connected simultaneously

The connectors panel moment: that list is your attack surface

Glasses ON: enterprise governance gap

Glasses OFF: the Minecraft PC Standard still applies

Try This Yourself: n8n, one test workflow, 30 minutes

Read the full article: https://open.substack.com/pub/aifrankly/p/mcp-103-i-gave-claude-code-someone

Watch on YouTube: https://youtu.be/0idx_l2o5X0

Settings:

Title: MCP 103: I Gave Claude Code Someone Else's Tools (Full Deep Dive)

Season 1, Episode 11, Full

RSS toggle: ON

Send via email: OFF

No paywall

Ready to upload the audio?