Most headlines read this as a story about hackers. I read it as a deadline for small teams.
On Monday, Google's Threat Intelligence Group said it found a framework called exploit_agent. VentureBeat reported that it compares software code before and after a patch, spots the flaw the patch fixed, drafts an exploit and runs it, with no human in the loop. Google says it is the first time it has seen attackers fully automate that chain. The Hacker News adds that Google also tracks other agent frameworks, including CAIRN and PentAGI, and points to PaperCut NG/MF flaws earlier this September as an example of newly disclosed bugs being weaponised fast.
Two limits are worth saying out loud. Google did not name a group or a country. And the report does not show that exploit_agent itself caused any specific break-in.
Here is my take. For years, a quiet advantage of a small team was boring. Turning a patch into a working exploit took a skilled person and time, so most small shops had a few days of cushion. Merritt Baer, a former deputy CISO at AWS, told VentureBeat a skilled person might need days for this work. That is her estimate, not a Google finding. If a machine does the same job, I would not plan around days.
The popular line says AI will help defenders more than attackers. Maybe it will. But a five-person team does not have a defender. It has whoever set up the tools.
So my rule has nothing to do with buying more AI. Know what you have exposed to the internet. Know who owns each tool, including every AI app, plugin and server someone stood up for a quick test. Patch the exposed ones first, within days, not at the end of the quarter. Baer's advice in the same article is close: know what you expose, patch faster, watch for exploitation.
One detail made me smile. The attacker's own setup file told the operator to paste LLM API keys straight into it. Do not copy that habit.
Patching used to be a chore. Now it is the race. Plan like it.
Written by
Kristjan SaarikTech entrepreneur building AI agents26 pieces · 11 followersFollow Kristjan to see new posts in your feed, on your phone or by email.
More from
Kristjan Saarik →Cursor told Congress it may "phase out" Chinese AI models
Cursor told two House committees it wants to "explore whether it can phase out" Chinese AI models, CNBC reported on Friday. Nothing changes for you today. It is a statement of intent, not a decision.…
"The most recent safe version is 1.26.0." Pin it today
Most headlines call this a Microsoft story. For a small team, I think it is a lockfile story. Wiz reported that two malicious versions of Microsoft's agent-framework package, 1.26.1 and 1.26.2, were…
United says no to AI agents booking flights. Why small teams care
United Airlines will block AI agents from booking flights for customers. The reason is not the technology. It is the password. TechCrunch's Zack Whittaker reported it on Thursday. United says it will res…