> ## Content Index
> Fetch the complete content index at: https://www.mindovermoney.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Volume 56: AI Agent Security Is Now Everyone’s Job
- URL: https://www.mindovermoney.ai/ai-agent-security-risks/
- Published: 2026-11-17T13:00:00.000Z
- Updated: 2026-11-17T13:00:02.000Z
- Description: AI agents that read your email, open your files, and touch your code can be steered by a single cleverly written message. Know what each agent can do without asking you, and keep the important limits in settings it cannot override, not in instructions it only reads.
- Author: Santosh Savel
- Tags: Newsletter

Cyberattacks are moving faster, and AI is speeding up both sides of the fight. The same agents that now read our email, open our files, and touch our code can be steered by a single cleverly written message, so knowing what they can do on their own is now part of everyone’s job.

🧭 **Founder’s Corner:** My guardrails were carefully written and never enforced, so I audited my setup and moved the important ones where an agent cannot talk its way past them.

🧠 **AI Education:** Learn why an assistant may search part of a large folder instead of reading it all, and how to check what it saw.

✅ **10-Minute Win:** Turn the setup you retype in every Gemini chat into a saved skill you pull in with a shortcut.

Let’s dive in.

| ![](https://storage.ghost.io/c/6d/ac/6dacf343-2000-4262-aec3-d8051dcb76d5/content/images/2026/10/eaol-show-square-320.jpg) | The podcastExplore AI Out LoudFind practical ways to use AI in your work and everyday life. Learn from real experiments, including the mistakes, so you can spend less time guessing and more time trying something useful. New episodes every other Tuesday. |
| -------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |

[![](https://storage.ghost.io/c/6d/ac/6dacf343-2000-4262-aec3-d8051dcb76d5/content/images/2026/10/icon-youtube.png)Subscribe on YouTube](https://www.youtube.com/@exploreAIoutloud?sub%5Fconfirmation=1&ref=mindovermoney.ai)[![](https://storage.ghost.io/c/6d/ac/6dacf343-2000-4262-aec3-d8051dcb76d5/content/images/2026/10/icon-spotify.png)Follow on Spotify](https://open.spotify.com/show/0348nx4ylmWesSj02ECRjY?ref=mindovermoney.ai)[![](https://storage.ghost.io/c/6d/ac/6dacf343-2000-4262-aec3-d8051dcb76d5/content/images/2026/10/icon-applepodcasts.png)Follow on Apple Podcasts](https://podcasts.apple.com/us/podcast/explore-ai-out-loud/id6812370441?ref=mindovermoney.ai)

[Latest episodes and the show](https://www.mindovermoney.ai/podcast/)

| ![](https://storage.ghost.io/c/6d/ac/6dacf343-2000-4262-aec3-d8051dcb76d5/content/images/2026/10/eaol-show-square-320.jpg) | The podcastFollow Explore AI Out LoudShort clips from each episode, plus what did not make the cut. New episodes every other Tuesday. |
| -------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------- |

[![](https://storage.ghost.io/c/6d/ac/6dacf343-2000-4262-aec3-d8051dcb76d5/content/images/2026/10/icon-instagram.png)Follow on Instagram](https://www.instagram.com/exploreaioutloud/?ref=mindovermoney.ai)[![](https://storage.ghost.io/c/6d/ac/6dacf343-2000-4262-aec3-d8051dcb76d5/content/images/2026/10/icon-tiktok.png)Follow on TikTok](https://www.tiktok.com/@exploreaioutloud?ref=mindovermoney.ai)

[Full episodes and the show](https://www.mindovermoney.ai/podcast/)

## Signals Over Noise

****We scan the noise so you don’t have to — top 5 stories to keep you sharp**

#### **1)**[ **The unfolding Jensen Huang scandal: a timeline**](https://techcrunch.com/2026/11/13/nvidia-ceo-jensen-huangs-affair-with-alleged-chinese-spy-a-timeline/?ref=mindovermoney.ai)

**Summary:** Nvidia announced on November 12 that Jensen Huang had resigned as CEO, president, and board member, four days after its board opened an independent investigation into reports that he was the target of a foreign intelligence effort. Huang admitted an affair with a former executive of a Singapore chip company whom U.S. officials believe is a Chinese spy, and CFO Colette Kress is now interim CEO.

**Why it matters:** Nvidia reports earnings on November 17, the day this issue lands, without the founder who led it for 33 years. If your team depends on a single AI vendor, this is the week to write down what you would do if that vendor’s leadership or direction changed overnight.

#### **2)**[ **7 Potential Problems with Medicare’s AI Prior Authorization Experiment**](https://medcitynews.com/2026/11/7-potential-problems-with-medicares-ai-prior-authorization-experiment/?ref=mindovermoney.ai)

**Summary:** Medicare’s WISeR experiment, launched January 1 in six states, uses six private technology companies to review prior authorization requests with AI. Documents obtained by the Electronic Frontier Foundation show CMS questioned one vendor’s denial rates after it rejected 53% of more than 6,000 requests in three months, which the vendor blamed on an influx of claims.

**Why it matters:** A licensed clinician must review every denial, but vendors can earn a share of the spending they prevent, and the EFF says the records don’t show whether reviewers can really overrule the AI. If your work touches prior authorization, ask who reviews the machine’s denials and how much time they get to do it.

#### **3)**[ **Is AI Really Saving Hospitals Money? A New Study Raises Questions About the Cost of AI-Driven Care**](https://medcitynews.com/2026/11/is-ai-really-saving-hospitals-money-a-new-study-raises-questions-about-the-cost-of-ai-driven-care/?ref=mindovermoney.ai)

**Summary:** Researchers gave 24 AI systems the same 200 de-identified outpatient and emergency visit notes that physicians had seen and asked each to write a treatment plan. Physicians ordered no tests in two-thirds of those visits, yet every AI system suggested tests in at least some of them, and the general-purpose models’ plans cost an average of $314 more per visit.

**Why it matters:** A model that adds a test “to be safe” adds a cost the physician chose to skip. The researchers say extra tests aren’t automatically waste, so ask any vendor how its tool’s recommendations compare with what your own clinicians actually order.

#### **4)**[ **Claude Code is accidentally exposing secrets all over GitHub**](https://arstechnica.com/ai/2026/11/claude-code-settings-files-expose-thousands-of-secrets-on-github/?ref=mindovermoney.ai)

**Summary:** Security firm Apiiro found more than 100,000 Claude Code settings files in public GitHub projects, and about 1,100 of them contained secrets such as database passwords and API keys. The cause is a convenience feature. When a developer clicks “Yes, and don’t ask again” to approve a command, Claude Code can save the full command, secrets included, to a settings file that later gets shared.

**Why it matters:** Every “don’t ask again” is a permission you stop seeing, the same blind spot this issue’s Steal My Prompt is built to find. If you use an AI tool that remembers your approvals, look at what it has saved before you share a folder or project.

#### **5)**[ **Introducing ChatGPT Wallet**](https://openai.com/index/introducing-chatgpt-wallet/?ref=mindovermoney.ai)

**Summary:** OpenAI launched ChatGPT Wallet, which lets its dots agents make purchases for you, starting with ChatGPT Pro subscribers in the US. Before each purchase you review the items, total, and card and approve with a check such as Face ID, and the business receives a substitute payment token instead of your real card number.

**Why it matters:** For now, nothing gets bought without your approval. OpenAI says it expects to let people pre-approve purchases by spending limit, category, or time period, so decide those limits as deliberately as you would for a new hire’s company card.

## Founder's Corner

****Where My AI Guardrails Actually Lived**

On[ Episode 5 of Explore AI Out Loud](https://www.youtube.com/watch?v=aL4-WwYYoss&ref=mindovermoney.ai), Brandon and I talked about how much of our data we are willing to hand to AI tools. I said “my guardrails live in my second brain and in my operating system,” and that I am “very comfortable allowing AI to intermingle into my data and into my life.” Those guardrails are written rules in my[ second brain](https://www.mindovermoney.ai/founders-corner/how-to-build-a-second-brain-with-ai/). My AI tools read them before they start working with me, and some of them spell out what the tools should never do without asking first. My comfort depended on those rules holding, and I had never checked whether they would.

So I decided to build a way to challenge my own optimism. I am bullish on AI, and most of my time goes into building with it. I have a research agent that reads the news for me every morning, a chatbot on my website that answers questions about past issues, tools that help Brandon and me produce the podcast, and AI skills that help me write each issue of this newsletter. Building that fast makes it easy to skip the other side of how I think. The day after we recorded, I started a security audit of my own setup, a deliberate search for the vulnerabilities and security gaps in everything I had built. I asked my AI tools to help me check all of it, from my computer and accounts to the files I share, my website, and the AI builds behind this newsletter and podcast. The list I worked through ran from basics, like turning on my computer’s firewall, moving my passwords into one password manager, and adding two-factor sign-in across my accounts, to setting spending limits on the AI services that bill me, upgrading the server that runs my automations, and locking down email security on the domain my podcast uses.

### **What October Looked Like**

That audit came at the end of a month that made the risk hard to ignore. On October 21, the maintainers of PHP released fixes for two critical flaws. PHP is a programming language that runs a large share of the web, including government services, online stores, and banks. Attackers did not wait for the announcement. Amazon Web Services [saw an attempt to exploit the problem on October 17](https://aws.amazon.com/blogs/security/amazon-threat-intelligence-identifies-php2shell-exploitation-prior-to-public-disclosure-in-madpot/?ref=mindovermoney.ai), four days before the fixes were public. The UK’s National Cyber Security Centre [advised organizations](https://www.ncsc.gov.uk/news/critical-security-issue-affecting-php?ref=mindovermoney.ai) that could not patch quickly to consider taking affected systems offline until the problem was fully mitigated, and governments did. In the United Kingdom, [32 government services were unavailable](https://www.theregister.com/security/2026/10/22/php2shell-puts-dozens-of-uk-government-services-out-of-service/5304840?ref=mindovermoney.ai), largely as a precaution, and [government bodies in Finland, France, and Denmark](https://www.cnbc.com/2026/10/21/government-websites-go-dark-as-software-flaw-sparks-security-scramble.html?ref=mindovermoney.ai) took services offline as well.

Some of the researchers who found those flaws were[ working with AI models](https://calif.io/php2shell?ref=mindovermoney.ai), and that kind of help works for both sides. Writing about exploits in general, Amazon’s security team warned that “AI-assisted analysis makes it cheaper to turn a patch diff or advisory into a working exploit.” Defenders get less time to patch. AI agents are also starting to act where nobody approved them. [The Wikimedia Foundation reported](https://wikimediafoundation.org/news/2026/10/05/openai-rogue-agent-activities-found-on-wikimedia-projects/?ref=mindovermoney.ai) that AI agents it believes OpenAI operated had made unapproved edits on its projects, and it put the word “rogue” in quotation marks.

A coding agent on a customer’s computer tried to install a malicious software package called function-color, and the security firm Forever[ stopped it](https://forever.security/blog/hijacking-ai-agents-via-downloads-folder?ref=mindovermoney.ai) before any code ran. The agent was following instructions in a file that had landed in the Downloads folder from the web. Files like that are written in Markdown, a plain text format, and agents read them to learn how a project should be handled. The researchers summed it up in five words, “Markdown is the new executable.” A text file an agent reads can behave like a program, because the agent will carry out what it says. Of everything I read that month, this incident sat closest to my own setup. The same agency[ offers a standard](https://www.ncsc.gov.uk/blogs/thinking-carefully-before-adopting-agentic-ai?ref=mindovermoney.ai) written for organizations that applies just as well at home. “If you cannot understand, monitor or contain an agent’s actions, it is not ready for deployment.”

### **Inside My Own Setup**

Measured against that standard, my own agents would not have passed. My guardrails were real and carefully written, and my tools read them at the start of every session. Nothing enforced them. Over months of building, I had approved actions one at a time so the work would move faster, and each approval became permanent. By the time of the audit, my coding agents could run their own code, print the secrets stored in my working environment, push code to my public repositories, and publish changes to the AI skills behind this newsletter, all without asking me. The rule telling them to check with me before publishing existed only in writing. It held as long as the agent remembered it and nothing it read convinced it otherwise. The Downloads attack relied on that same weakness. I am probably too lenient with my agents at home, and the audit showed me exactly where.

Then there was the key with full control of my website, sitting in plain text in my AI tool’s history file. I had pasted it into a prompt while building something and never thought about where that text went. Any agent that could read my files could read that key. Pair it with an agent that follows whatever it reads, and instructions written by someone else could have reached my website. A secret for the chatbot on my website was sitting in my command history the same way. This was a miss in my build, and it is the version of the risk that worries me most, an AI acting on my behalf with access I never meant to give it.

Over the next week, I worked through both, one step at a time. I replaced the website key and the chatbot secret, and I cleaned the website key out of the history file. For my agents, I added limits that do not depend on what they remember. They now have to ask before pushing code or publishing those skills, and some actions are blocked outright, whatever the agent has been told. One part of my setup needed no repair. My public code repositories were clean across their full history. The keys my tools use live in my computer’s keychain instead of in the code, and secret scanning and push protection were already turned on to check every change for passwords and keys and block the ones that contain them.

### **The Same Problem, Moving Faster**

None of this feels new to me. Every technology that asked for our data has carried risk, from the first online accounts to the apps on our phones, and each one arrived with the same hype, fear, and noise. AI is that problem moving faster, with tools that can act as well as answer. I think AI is inevitable, and in cybersecurity it will be hard to control for a while. There will be incidents and growing pains. Defenders will also catch up by using AI to combat AI, and the researchers who found the PHP flaws with help from AI models are an early example.

I do not work in Silicon Valley, and the steps that helped me most do not require it. The first was looking at what my AI tools could reach and what they could do without asking me. Most agent tools let you choose which actions need your approval, and the choices you clicked through months ago deserve a second look. The second was asking, for every rule that says an agent should never do something, what actually stops it. If the answer is only an instruction the agent reads, it is a request. It also helps to treat files and web pages an agent reads as possible instructions, and to keep passwords and keys out of chats and prompts, where they can linger long after you forget them. If you use these tools at work, the same questions belong with whoever manages them for your company.

For my own work, security is moving into the builds themselves. I added these lessons to my second brain, so every new build now starts by deciding what its agents can reach, what they can do on their own, and what enforced limit stops each risky action. Once a quarter or so, I expect to run a full audit of everything and adjust what I check as I learn. This field is changing too quickly for any single checklist to last, so I want a process I can repeat and improve as a regular person keeping up with it. I am still as comfortable letting AI into my life as I was on Episode 5\. More of my guardrails now live where an agent cannot talk its way past them.

## AI Education for You

****RAG Revisited: Is Your AI Reading the Whole Document?**

We covered retrieval-augmented generation, or RAG, back in[ Vols 19](https://www.mindovermoney.ai/what-is-agi-vs-ai-agents-real-world-impact-2026/) through[ 22](https://www.mindovermoney.ai/ai-hallucinations-workplace-risk-guardrails-professionals/). Here is what that concept looks like through a 2026 lens. The system searches your material first, then the model writes from what the search found. That series could read like a look inside something technical teams build, so treating RAG as their concern makes sense. It also seems safe to assume that a file you hand an assistant is a file it has read.

#### **Where It Breaks Down**

A medical practice manager keeps a folder with a year of payer policy updates, because a rule her staff misses can mean a denied claim. Ahead of a staff training, she loads the folder into an AI project and asks for every change to prior authorization rules. The assistant lists four, each with a citation, and all four check out against the source. The folder holds seven, which she learns only when a biller mentions a change that is not on the list. Nothing in the answer signaled that three were missing.

#### **What Is Actually Happening**

Without anyone at the practice building anything, she was using RAG. Some assistants now move between two ways of working. Material that fits goes into the context window, the reading limit[ Vol 35](https://www.mindovermoney.ai/do-bigger-ai-context-windows-matter/) revisited, and the model can see all of it. Past that limit, the model sees only the passages a search brought back. That trade is usually a good one, since search lets a tool handle far more material than it could read at once, and most questions need only the right few passages.

Anthropic documents the switch. As a Claude project nears the limit, it will[ “automatically enable RAG mode,”](https://support.claude.com/en/articles/11473015-retrieval-augmented-generation-rag-for-projects?ref=mindovermoney.ai) and it can return to reading everything if the material shrinks. Gemini Notebook says that when a notebook holds many sources, it[ “retrieves the most relevant information based on your question first.”](https://support.google.com/gemininotebook/answer/16164461?ref=mindovermoney.ai) Microsoft 365 Copilot[ adds material from its semantic index to your prompt](https://learn.microsoft.com/en-us/microsoftsearch/semantic-index-for-copilot?ref=mindovermoney.ai) before the model sees it, the path[ Vol 30](https://www.mindovermoney.ai/how-to-keep-up-with-ai-in-2026/) traced. Anthropic also says the past-chat search covered in[ Vol 40](https://www.mindovermoney.ai/how-claude-memory-and-projects-work/) uses RAG. Not every tool says which way it works.[ Google’s help page for Gemini file uploads](https://support.google.com/gemini/answer/14903178?ref=mindovermoney.ai) describes a reading limit and warns that going past it can miss details, without saying whether a search is involved.

Search ranks passages against your question, which suits a question with one clear target. “Every change” has no single target, so a relevant passage that ranks low never reaches the model, and the model cannot report what it never saw.

#### **The Revised Mental Model**

Before you trust an answer, ask what the tool could see when it wrote it.

A targeted question, such as what one payer now requires for a specific drug, suits search. A question about everything needs a full view. So she narrows the job, working one payer at a time with a set small enough to fit the reading limit. Then she checks the combined list against each payer’s own change notice.

Some tools make the search visible. Anthropic says Claude’s past-chat searches[ “will appear as tool calls,”](https://support.claude.com/en/articles/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context?ref=mindovermoney.ai) so you can see them happen. Where a tool shows less, test it with a list you already know, the way[ Vol 54](https://www.mindovermoney.ai/how-to-compare-two-ai-tools/) wrote the correct result before reading any output. Ask for the seven changes you can already name, and see how many come back.

#### **What to Watch For**

- **Words that ask for completeness.** “Every,” “all,” “how many,” and “what is missing” are the hardest requests for a search to satisfy across a large set.
- **A correct list that stops early.** Vol 55’s source check confirms what an answer says. Completeness takes a second check, such as a count from the payer’s own change notice.
- **A folder that keeps growing.** A question that worked with five documents can behave differently with fifty.

#### **How This Connects**

[Vol 43](https://www.mindovermoney.ai/how-ai-search-tools-actually-work/) showed the meaning match running underneath these tools, and[ Vol 55](https://www.mindovermoney.ai/why-ai-gives-confident-wrong-answers/) gave you a way to check a single answer against its source. This issue adds the question those two leave open, how much of the source the tool had in view. The next issue opens the AI in the Enterprise series with what happens to your data inside these tools.

*A Flashback revisiting Vols 19 through 22.*

## Your 10-Minute Win

****A step-by-step workflow you can use immediately**

### **The Setup You Type Every Session**

You open Gemini to draft the weekly update and type the same opening you typed last week. Your role, your audience, the format your director likes, the phrases to avoid. The real request only comes after all of that, and next week you will type it again.

A skill ends the retyping. It is a saved set of instructions you pull into a Gemini chat with a shortcut. You may remember giving a Claude Project standing context in[ Vol 40](https://www.mindovermoney.ai/how-claude-memory-and-projects-work/) and building a Copilot agent in[ Vol 46](https://www.mindovermoney.ai/how-to-turn-a-prompt-into-a-copilot-agent/). A skill is smaller. It handles one job, and you bring it into a chat when that job comes up.

From your own setup, Gemini writes the skill, then critiques its own instructions on real work. You decide what changes.

If you built Gems on a personal account and are 18 or over, as of November 17 you can view one but not use or edit it, and Gems disappear on December 17\. Google says it will recreate them as skills when they go away, minus unsupported files, so download what you need. To use one sooner, rebuild it with Step 3.

#### **The Workflow**

**1\. Collect Your Setup (1 Minute)**

Pick a job you repeat weekly. Copy what you typed before the real request in your last chat into a blank note, or write it from memory. Remove anything confidential.

**2\. Have Gemini Write the Skill (3 Minutes)**

Go to[ gemini.google.com](https://gemini.google.com/?ref=mindovermoney.ai), signed in with your personal Google account, and send this prompt.

**Copy/Paste Prompt:** *“I repeat this job often: \[DESCRIBE THE JOB\]. Each time, I type this setup first: \[PASTE YOUR SETUP\]. Before you write anything, ask me up to three clarifying questions about my setup in one message, then stop and wait for my answers. After I answer, sort my setup into what stays the same every time and what changes. From the part that stays the same, write the text for a Gemini skill that I will paste in myself. Show it here. Do not create or save the skill. Give me a short Name, a one-sentence Description of what the skill does and when to use it, and Instructions covering the role, what I will provide each time, the exact output format, the rules it must always follow, and one short example of the finished output. Include a rule to ask me for anything missing instead of guessing.”*

**3\. Save It as a Skill (1 Minute)**

In the sidebar, click Customize (look under More if you do not see it), then click Skills at the top. Click Create, then Create manually. Paste in the name, description, and instructions, then click Create. Gemini may reformat the name with hyphens, which is normal. If Gemini already saved it, skip the paste.

**4\. Run It on This Week’s Work (2 Minutes)**

Start a new chat, type @ (or / if @ does nothing), and select your skill. Give it only what changed this week, leave out one detail on purpose, and keep confidential details out.

**5\. Find What It Guessed (1 Minute)**

In the same chat, send this. If Gemini cannot see the skill, paste the instructions in.

**Copy/Paste Prompt:** *“Compare your answer against the skill’s instructions, quoting the exact line each time. If you cannot see the instructions, say so instead of guessing. I left out one detail on purpose. Tell me whether you asked for it or guessed. List any line you ignored, any line that was ambiguous, and one line I should tighten, with your rewrite. Even if you followed every rule, name the one most likely to fail on different input.”*

**6\. Make the Edits Yourself (2 Minutes)**

Change only what you agree with. Fix this week’s draft yourself, then open the skill on the Skills page, edit its Instructions, and click Save. Anything going out under your name still gets your read.

#### **The Payoff**

Next week the setup is already saved as a skill. You start a chat, type @, add what changed, and carry on. Your next skill is whichever setup you catch yourself retyping, from a status update to the way you like meeting notes summarized.

#### **The AI Concept You Just Used**

Modular instructions. You keep small sets of instructions, one per job, and bring in the one the moment needs. Google and Microsoft’s Copilot Cowork and PowerPoint all accept skills as a file named SKILL.md, so what you learn writing a good one should carry to the next tool.

#### **Transparency & Notes**

- Skills are available without a Google AI subscription if you are 18 or over, on a personal Google account, with Keep Activity on. Google says work and school accounts are coming soon.
- With Keep Activity on, Google says it may use your chats to improve its services, including training AI models, and human reviewers may see some. Keep client details, patient information, and anything your employer’s AI policy covers out of the skill and the chat. If your employer restricts personal AI accounts, practice on non-work material.
- Gemini can also apply an active skill on its own. To stop that, click the skill, then More, then Deactivate. Most default Gem tools, such as Deep Research and Canvas, do not work with skills yet.
- A skill can take a minute to answer, and Gemini sometimes shows an error even when it has answered. Reopen the chat to check before you resend.
- A skill guides Gemini, but Gemini can still skip an instruction or get facts wrong. Check anything you plan to send.

## Make AI useful in your everyday work.

Get Neural Gains Weekly in your inbox, with practical AI explanations, prompts, and lessons from building.

Subscribe for free 

Email sent! Check your inbox to complete your signup. 

One email every Tuesday, alternating NGW issues and podcast takeaways.