A short update from my corner of the internet.
You may have noticed it’s been about two months since you’ve heard from me (newsletter-wise).
Admittedly, I’ve been trying to figure some things out since I recently quit my full time job (yes I’m soft launching it here. Ya’ll heard it first). I even took a month off to travel Europe and attend Cannes Lions, which I’m sure you heard plenty about if you follow me on IG or LinkedIn.
You’re probably wondering if I quit to pursue content creation full time, but the answer is absolutely not. More details on that to come, but quite frankly, it’s impossible to post content about AI Security without being directly involved in the industry since things are changing so fast.
Traditionally on Substack I was only sharing weekly AI, Security, and Quantum news I found interesting. Starting this week I will also be sharing AI tools I’ve tried, random things I’ve been thinking about, AI workflows I’ve been building, or some fun things I have been up to. So you can expect to see a good mix of things!
What caught my attention.
01/ First documented agentic ransomware by an Agentic Threat Actor (ATA)
I came across the term ATA for the very first time in this writeup documenting the first agentic ransomware, which is a fully AI agent powered attack that uses self narrating code and “machine speed” error debugging.
The main takeaway here is old vulnerabilities you risk accepted or ignored in the past may come back to haunt you because AI makes it trivial to enumerate and exploit old vulnerabilities at scales. Which also means you should be really diligent about patching any third-party dependencies.
02/ Google account takeover phishing campaign targeting job seekers
For all those looking for new jobs or like clicking on links sent from recruiters, this months’ long campaign impersonates big name companies like OpenAI and Netflix to steal your Google credentials.
I always find campaigns like this scarier than more technical AI vulnerabilities because they target the masses. The only way you can protect yourself against this is to be really diligent about checking the email addresses and links before you click on them, and especially before logging in with your credentials.
I recorded a video explaining in more detail so do check it out.
Watch the video →
03/ Parasitic supply chain attack impacting CI/CD workflows
A new supply chain attack called Cordyceps is giving HBO Max hit The Last of Us vibes. It lets malicious pull requests take over privileged workflows, meaning an attacker with a free GitHub account could potentially steal credentials, push code, and quietly make their way into the software supply chain.
04/ Vulnerability in GitHub’s new agentic workflows
Along the lines of workflows, this GitLost vulnerability exploits GitHub’s new agentic workflows, which let you describe workflows in natural language. Unsurprisingly, the vulnerability leverages a pretty classic form of prompt injection by exploiting the agent’s context window.
05/ GhostCommit in the code reviews
This is a pretty genius form of indirect prompt injection where the payload is delivered via photos because LLM and human code reviewers are only reviewing text. Seems like harness is especially important here because the same model will both flag or miss the exploits.
Use passkeys where you can.
SMS codes are better than nothing, but they’re one of the weaker forms of MFA. If an account lets you set up a passkey, do it. You can sign in with Face ID, your fingerprint, or your device PIN instead of relying on a code sent to your phone.
AI is not a magical potion.
My pet peeve is seeing people post things like “I built this website in 5 minutes” or “AI helps me make $10k+ a month” because it gives the false sense that you can one-shot prompt your way to success. For example, I built a website using both Codex and Claude thinking it would take me only a few hours. It took me THREE days. Check out what worked, didn’t work, and security concerns I had while building it below ⬇️
Watch the video →
I got an iPad recently and have been obsessed with three things:
- using it to remote into Codex on my laptop
- watching cdramas on it
- using my first iPad app built fully with the Codex agent Xcode beta. It also helped me get it out to TestFlight 👇🏻
I stumbled upon Tadaima in SF the other day and can’t believe I haven’t had their matcha before. The strawberry cream matcha was really good. Actually flavorful and strong matcha.
That’s all I have for this week. If you have questions, comments, or feedback, feel free to reply directly.
If this newsletter was useful, forward it to someone who might find it helpful too.