Uncomfortable facts: Human code is brittle. Coding agents have made code abundant and cheap. Open models keep getting better; and incidents where agents exploit vulnerabilities in APIs and Frontends written by humans have happened and will increase in volume as tokens get smarter and more accessible.
Black Hat USA 2026 | The 'Breaking' News: The OpenAI–Hugging Face Incident
This poses a huge security risk, in particular for companies with limited budgets.
But there are things we can do to proactively tackle the problem and try to reduce the risks of being hacked by remote agents.
This is what we will be implementing at Cubiq in the next few weeks, of course in our bootstrapped but efficient way.
Agentic code auditing
We will assume that there are two types of code audits: Statics and Runtime audits.
In this article we will talk about the former.
At Cubiq we use Opencode as our harness for synchronous coding assistance, shoutout to Kimi K3 Max, Qwen 3.8 max and GLM-3.8-Flash. But Opencode could be used for async tasks like periodic, automated, multiple-agent static code analysis and auditing.
We started out by writing a simple bash script with a prompt to test this out against our Cubiq API.
We asked the agent to order findings by their severity and to notify the summary to our company Telegram group.
Scary stuff that has all been fixed by now, but still… Cubiq has several APIs, from E-Commerce to Logistics. And these low hanging fruits were well... Hanging in there, waiting for the catch. Like a "coldcard" bug 😉
You can find the code here.
We needed to make this better by instructing the agent to search security vulnerabilities databases like CVE https://www.cve.org/ and the kind; and apply the findings to the audit.
Then we decided to use the Cloudflare's security-audit-skill as a base structure for our script, enhancing the prompt to also look for high-impact improvements within the code.
As any bash script it can be used for automation; the whole idea behind our audit repo is to receive a Telegram notification after each security audit (they compound!) automatically fired via a cronjob in regular intervals. But the Cloudflare skill can be used on its own, interactively via Opencode TUI as well.
Using the script you will end up with a .audit-results folder in your projects root directory.
Use it, you can have some fun doing some fixing.
So, rogue agentic attacks are not a thing of the future or a cool Sci-Fi Cyberpunk anime movie.
When AI agents write the code, and autonomous agents scan the web to exploit it, your defense must become one too. If you aren't using local agents to hack your own stack, rogue agents across the wire will do it for you.
IT companies, and more broadly every project exposed to the Internet will need to step up, be aware of the new risks agentic LLMs pose, a new threat model and a new speed at which things can go wrong and take countermeasures now.
This of course might become an industry in itself. One thing is for sure, we are not going back, even after the token-spending fever stabilizes.
Game is on!
I explore more in-depth ideas about agentic code auditing in some Nostr article (English only).
(Don't know what Nostr is? Check it out).
Be safe!
