It’s been an interesting couple of months in security. Last month was “Black October.” There were a number of supply chain attacks, including one on LiteLLM, for example, its second compromise this year and the first of which I wrote about earlier when it happened. It’s like déjà vu all over again.
Anyway, a week prior to that, we saw the ConnectWise compromise, which Microsoft later attributed to a Chinese nation-state group using ARTEX, an open-source, multi-agent AI penetration-testing framework that attackers had weaponized.
Then you had these bangers. Critical vulnerabilities turned up in PHP and Node.js, which exposed a pretty large share of the internet to attackers. The PHP flaws had been dormant in the code for decades, and it took frontier AI to chain together the steps needed to exploit them.
The NSA’s advice to organizations that couldn’t patch right away was to take your systems offline. You know things are wild when a national intelligence agency’s fallback advice is “unplug it.” Extraordinary.
In software development, engineers are rethinking almost every assumption about how we build software because of AI. Some are deleting unit tests, forking all their dependencies, pushing straight to main, and even writing their own version control. It’s a time of rapid reappraisal for pretty much everything in our industry.
I believe security is going through that same kind of seismic shift right now, too.
AI has blown past the speed at which human security researchers can work. This situation is exactly what Anthropic warned about back in April, when it announced Project Glasswing and Claude Mythos Preview, a model it considered too dangerous for public release.
At that point, Mythos had already found bugs like a 27-year-old flaw in OpenBSD and a 17-year-old remote code execution bug in FreeBSD. I recall people for whatever reason handwaving at the time, believing this was some type of doomer marketing, but for anyone paying attention it was clear we were on the cusp of a new era.
The toothpaste wasn’t going back in the tube.
Serving the slop
And there is another problem that I hadn’t anticipated though, which is slop CVEs. Maintainers are inundated with a deluge of them, and not all may apply to your deployment depending on your code paths and environment.
But you still gotta figure it out and it takes time. I’ve seen a single Node.js bug turn into dozens of hours of unplanned work for one team. The thing you need is proof, which will absolutely determine the actual urgency.
That’s where I believe an always-on agent comes in, the kind with its own computer. Since OpenClaw earlier this year, that class of agent has become widely available for normies in products like Claude Duet, Dots, Grok Bot, and Muse. What I’m describing here is the same kind of agent, but focused entirely on security.
A security agent can spin up as many parallel threads as it needs and operate 24x7, guarding your actual deployment. We have to fight fire with fire, though I believe the devil will be in the details. You’ll preferably need access to cyber models like the ones offered through OpenAI’s Daybreak or Anthropic’s Project Glasswing.
That said, I mentioned ARTEX earlier. It is curious and notable to point out that while ARTEX can operate with a frontier model, attackers don’t even need frontier or special cyber models to do damage. In a campaign against South Korean financial firms, for example, CrowdStrike found the attackers ran ARTEX on DeepSeek-V4.1-Flash as the “primary LLM backend,” supplemented by GLM-5.3 and Grok 4.6.
What I’m getting at here is that the best open weight models today already meet attackers’ needs. The flip side is that while I’d recommend access to cyber models from Anthropic and OpenAI for defense and pentesting as mentioned, you can likely go far with open weight models too, especially for triaging findings and proving whether they’re exploitable in your own environment.
Again, my preference is for the locked down cyber models offered by OpenAI and Anthropic, but I’m going to be honest with you here. The actual track record of them being available in an emergency hasn’t been great.
When Hugging Face was breached in July by what turned out to be one of OpenAI’s own models, they found that commercial frontier models refused to analyze the logs. The company wrote that the guardrails couldn’t tell an incident responder from an attacker, so they ended up running forensics with GLM-5.2.
In October, OpenCode faced similar challenges during its breach. They were granted access to Daybreak in the middle of the incident, which I’m sure came in clutch. The problem here is you must have access to the models first (you should secure that now), but access to them can inadvertently be pulled in some unusual circumstances.
And you may recall this summer Anthropic’s most capable models were briefly pulled from everyone under export control. In other words, when you’re under attack, you don’t want your defense waiting on someone else’s approval. I’d always have a backup plan ready to go and that is open weight models if you have the GPUs.
Two worlds
We’re dealing with two worlds here, by the way. Before your code is deployed, you have time. You need to get features and bug fixes deployed, but you have time to do it safely.
There are already tools from the labs like Claude Security, Codex Security, Devin Security, or Google CodeMender. They scan your codebase, verify findings, and provide patches for you to review. That is a pretty key element here is that it sees a finding all the way through to a patch, not another one to just add to the pile.
These are all cloud offerings, so if you’re operating in more secure environments, you’ll have to roll your own though it isn’t rocket surgery, either. The main thing you need is access to the right models.
Every merge request kicks off a pipeline job where the agent reviews the diff and your dependencies against current CVEs. It then traces whether the vulnerable code paths are actually reachable and tries to exploit them in a sandbox. If an exploit works, the merge is blocked and a patch goes up for review. If it doesn’t, the change goes on its merry way.
The other side of the coin is your live deployments. That is to say the servers, containers, and services actually running in production, plus everything they expose to the outside world. And there is no luxury of time here.
Attackers were already trying to exploit the PHP flaws four days before a patch existed. Your always on agent knows what properties you own and what runtime versions are inside every container. When a new CVE releases, it can tell you within moments whether you’re exposed.
An always on agent also keeps probing your live surfaces for things no advisory will ever cover. That could be a misconfiguration, an oversight, or a genuinely new flaw. You want to find it before your adversary does.
For live deployments, your agentic security workers should follow the same rules of proof and patching. The difference is that here, you really are on the clock. Adversaries now move at machine speed, and I think it’s an easy prediction that the window of exposure will keep shrinking, from days to hours to seconds.
If the agent finds something critical, you may need to contain it first to give time to finish the proof. That might mean blocking a route, rotating credentials, or in the worst case taking a service offline.
How far you go should match what is at stake. For some systems, an outage is far cheaper than a breach. For others, a lighter touch will do.
Trust the model
I’d like to admit we need a human in the loop, but I’m going to say something that won’t be popular. I believe human in the loop is already cooked here.
It took me a few years to accept it, but I believe we’re eventually going to have to trust the model. Adversaries already operate at machine speed, and no human can make decisions that fast. You can’t have someone on standby, waiting to press a button every time a threat shows up.
That also doesn’t mean humans are out of the picture either. You’re the one designing the agent, so you decide ahead of time what it’s allowed to do on its own and where its limits are. But even then, the model is the one making the call in the moment.
You can give it all the instructions you want, and at the end of the day, it will do what it does. That’s the part we have to trust. I don’t know exactly when, but I predict it will become generally accepted that this is our only realistic option.
Something I’ve wondered about here as well is what role System One models could play in this scenario. In September, TypeSafe AI released Jev, a new model that doesn’t actually generate text. It returns a decision like yes or no or a choice from a list and with calibrated confidence, often in well under a second.
Pair a model like that with a capable cyber model, and you might get fast instinct for contain-or-not type decisions, and slower, deliberate reasoning for the proof and patch. I haven’t seen anyone try this for security yet, but it seems like exactly the right shape for the problem.
What’s at stake
The same machine speed attacks can reach the systems we all depend to live. Just this month, hackers remotely cranked up supermarket freezers across Europe. Stores in at least seven countries found refrigerators and freezers at 50°C (122°F), and had to throw out their food, weeks after a similar attack hit hundreds of supermarkets in France.
Water utilities have already had their controls tampered with. A ransomware attack on Colonial Pipeline caused fuel shortages on the East Coast. U.S. officials have warned that state-sponsored hackers maintain persistent access inside American critical infrastructure, including energy and water, waiting for the right moment to inflict harm.
Fuel moves the trucks, the trucks move food and medicine, and power runs pumps and the freezers. What really concerns me is several of these attacks landing at once, perhaps timed to coincide with a geopolitical crisis. Any of these on its own, we can recover from, but all of them at once and at machine speed keeps you up at night.
Closing thoughts
If I’m honest, you can see that I’m worried about the future. I’m worried about our infrastructure, and there are a lot of threats out there. But I believe we’ll solve this. I have to be optimistic, because that’s the only way I can get out of bed in the morning.
If you’re wondering where to start, I believe we as engineers must start early and start now. I would get access to cyber models before you need them. If you have access to GPUs, keep the most powerful, open weight abliterated model in your back pocket.
Whatever you’re responsible for at work, secure the codebase. Put a security agent in your pipeline first, then move on to your live deployments. Much of the work to be done is just good, old fashioned security work.
Our existence and survival are predicated on how we interface with machine intelligence.



