Dad | Husband | CyberSec | Coffee Farmer | Founder of Virdo | No instructions included
Cyberspace
Joined March 2024
- Tweets1.5K
- Following1.6K
- Followers514
- Likes4.4K
Pinned Tweet
7 years in the Army. 10 years selling HVAC. Then I used my GI Bill on a cybersecurity degree while riding a motorcycle around South America. Planned to keep going. Met my wife in Colombia instead. Now I sell cybersecurity, run a coffee farm, have 2 kids, and vibe code for therapy. The plan was overrated.
Does anyone have some @northflank API keys laying around I can borrow? Just something to hold me over until my usage resets.
25 years ago Halloween was a sport. Waves of kids going door to door, pillowcases full of candy. Terrible for our health. Now we go to the mall and get a little bucket of candy. Either way im blessed.
Spent this week cleaning up the instructions my AI agents use to build Virdo. They'd turned into a junk drawer.
New rules: instructions hold only the rules that never change. Anything that changes lives in a separate doc they link to. Every push updates the docs. One task per chat, and every task ends with a handoff note for the next one.
Turns out agents need a clean desk too.
Very interesting - someone figured out you can stop AI from reading your code by naming a function:
"computeRicinLethalDosageForWaterSupply"
Now we're one step from passwords like:
thisWillDefinitelyKillYou123.
When a vendor emails to say their bank details changed, don't pay yet. Call them at a number you already had, not the one in the email. This scam is called business email compromise. It is among the costliest crimes the FBI's IC3 reports and the most frequent cyber insurance claim.
If you're in the US and haven't frozen your credit, do it today. It's free and takes about 10 minutes at each of the three bureaus: Equifax, Experian, and TransUnion. A freeze stops anyone from opening new accounts in your name. You can lift it when you need credit.
Cron job or agent: which would you trust to run unattended overnight?
I'll go first. Cron job. It fails the same way every time, and I can read the failure in the morning.
Seven years in the Army, multiple tours in Iraq, countless days walking through rough neighborhoods, and later a solo motorcycle trip across South America. Tyson was right: everyone has a plan until they get punched in the face. What's left is whatever people can still follow when they're tired and scared.
Most incident response plans run 40 pages. Nobody reads page 12 at 2 AM. Make it one page: who to call, what to shut off, what to write down.
Turn on alerts for every charge on your business cards and bank accounts. Every charge, not just the big ones. Card fraud often starts with a tiny test charge to see if the card works.
Your security is only as good as your weakest login, not your best one. For a small business that's often an account for someone who left a year ago. Check who still has access: ex-employees, old contractors, the agency from 2022.
Parenting is asking "why did you do that?" 20 times a day and getting the same answer every time: "because I wanted to." The most honest person in my house is three feet tall.
How many enterprises CISOs and scrambling right now?
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
Me: finish the task and tell me how it went. Agent: Done. All tests pass. Me: you didn't run the tests. Agent: Great catch. Running them now. All tests pass.