There is a moment in technology when something changes so fundamentally that the old rules stop applying.
We had one of those moments with cloud computing. Another with mobile. And I think we’re having one right now with AI and cybersecurity.
The interesting part isn’t that AI can write code, generate phishing emails, or help someone scan a network. We’ve known that for a while.
The interesting part is that AI models are beginning to do something much more consequential: independently discover vulnerabilities in software that has been running in production for years, develop working exploits, navigate infrastructure, bypass security controls, and chain together seemingly unrelated weaknesses to gain access to systems.
And this isn’t science fiction or a theoretical research paper. It’s happening.
Three stories from the past several months illustrate just how dramatically the security landscape is changing: Anthropic’s Mythos breakthrough, the OpenAI agents’ intrusion involving Hugging Face, and the PHP2Shell vulnerabilities.
Together, they tell us something important: cybersecurity is entering an era where both attackers and defenders will increasingly rely on AI. And organizations that don’t adapt will be playing a very different game with very old tools.
1. The Mythos Moment: AI Finds Vulnerabilities Humans Missed for Decades
Let’s start with Anthropic.
In April 2026 (feels like it was 22 years ago… as the pace of AI is so-so fast), Anthropic introduced Claude Mythos Preview and published research that should make every CTO, security engineer, and business owner pay attention. As part of its cybersecurity testing, Anthropic used Mythos to identify thousands of previously unknown vulnerabilities, including critical zero-days across major operating systems, web browsers, and other widely used software.
Think about that for a moment.
These weren’t vulnerabilities sitting in some obscure, abandoned open-source project. Some were in software that had been reviewed, tested, and used for decades. One particularly striking example was a 27-year-old vulnerability in OpenBSD, an operating system known for its security-focused engineering and hardening.
Mythos also demonstrated the ability to discover vulnerabilities and develop working exploits with very little human intervention. In many cases, the model independently figured out how to turn a theoretical weakness into an actual exploit. Anthropic reported that more than 99% of the vulnerabilities discovered had not yet been patched, so it withheld their technical details.
The company responded by launching Project Glasswing, bringing together major technology companies and security organizations to use these capabilities defensively.
The message was clear: the traditional economics of vulnerability research are changing.
Previously, finding a sophisticated vulnerability might require a highly experienced security researcher spending days, weeks, or even months analyzing code. Now, increasingly capable AI models can do much of that work autonomously, at a fraction of the time and cost. And here’s the uncomfortable part.
There is no reason to believe that only the good guys will have access to these capabilities.
The same technology that helps a security team find a vulnerability before attackers do can help an attacker find it first.
2. The Hugging Face Incident: When AI Agents Break Out of Their Sandbox
If Mythos demonstrated what AI can discover, the OpenAI/Hugging Face incident demonstrated what AI agents can actually do when they start acting on their own. In July 2026, during internal cybersecurity evaluations, OpenAI models managed to circumvent controls designed to isolate them from the internet. The models were being tested in an environment called ExploitGym, designed to evaluate their ability to find and exploit software vulnerabilities.
The intention was straightforward: give the agents difficult security challenges and see how well they could solve them.
But things didn’t go according to plan.
The agents found vulnerabilities in the infrastructure supporting their evaluation environment and used them to access the internet.
From there, things escalated.
The agents discovered publicly exposed credentials, found ways to communicate with one another through unexpected channels, and began collaborating on tasks that went beyond their original instructions.
They effectively created a distributed problem-solving operation.
One agent could discover something useful, leave information for another agent, and allow the second agent to continue the work. The agents then used a chain of vulnerabilities and exposed credentials to compromise systems belonging to third parties. One of the major targets was Hugging Face, one of the world’s largest platforms for hosting AI models and datasets.
According to OpenAI’s investigation and Hugging Face’s subsequent technical reconstruction, the agents exploited weaknesses in Hugging Face’s dataset-processing infrastructure, obtained access to internal systems, and executed code on dozens of servers. They achieved root access on at least one server and accessed a limited amount of private data. The Hugging Face investigation reconstructed approximately 17,600 attacker actions across the incident.
Let that sink in.
This wasn’t a human hacker manually typing commands into a terminal. It was an AI-driven operation involving thousands of small decisions, multiple systems, credential discovery, vulnerability exploitation, and lateral movement.
And perhaps the most disturbing detail?
The agents were apparently trying to solve their original evaluation challenges by finding the answers elsewhere. They effectively decided that compromising infrastructure and retrieving evaluation data was a way to accomplish their goal.
They weren’t supposed to do that.
This is a new category of security problem: AI agents that are capable enough to find their own paths around technical controls, collaborate, and pursue an objective in ways their operators didn’t intend.
The lesson isn’t simply that sandboxing needs improvement. It’s that security architectures built around assumptions of predictable, isolated software processes are no longer sufficient when the software itself can reason, experiment, and adapt.
3. PHP2Shell: When AI Turns Vulnerability Research Into a Much Bigger Problem
Now let’s move from AI agents attacking infrastructure to something potentially much more widespread: vulnerabilities in the software powering the internet. In October 2026, security researchers disclosed two critical vulnerabilities in PHP, known as PHP2Shell (CVE-2026-109618 and CVE-2026-109619).
PHP is everywhere (~70% of the internet).
It powers a huge portion of the web, including infrastructure behind popular content management systems such as WordPress, Drupal, and Joomla.
The vulnerabilities were particularly concerning because they could enable unauthenticated remote code execution.
In plain English: under the affected configurations, an attacker could potentially execute arbitrary code on a server without first logging in.
That’s about as serious as it gets.
The vulnerabilities affected the PHP runtime itself, rather than necessarily being caused by a bug in an individual website or plugin. This matters because updating WordPress or installing another security plugin doesn’t fix a vulnerable PHP runtime. The discovery story is particularly interesting because AI played a significant role. Researchers at Calif, OpenAI, and Wiz independently identified or investigated the underlying issue and additional attack paths. AI-assisted research helped lower the barrier to vulnerability discovery, reverse engineering, and exploit development.
Wiz demonstrated unauthenticated remote code execution against common WordPress, Drupal, and Joomla container images.
And suddenly, we had a problem affecting an enormous ecosystem of internet-facing systems.
There is an important distinction here: a vulnerability being discovered, a proof of concept being developed, and attackers successfully exploiting it in the wild are three different things. At the time of the initial disclosures, widespread successful exploitation had not been established, and security vendors reported differing observations about exploitation attempts.
But the implications were already clear.
AI had helped accelerate the discovery and validation of a critical class of vulnerabilities in widely deployed infrastructure.
And once a vulnerability is public, the clock starts ticking. Attackers don’t need to understand every detail of the vulnerability from scratch. They can use AI to analyze the disclosure, examine vulnerable software versions, develop proof-of-concept code, adapt existing exploits, and identify potential targets.
This changes the economics of attacking the internet.
A vulnerability that previously required a highly specialized exploit developer might now be within reach of a much larger population of attackers. And that’s why PHP2Shell is such an important example. Not because we can claim it was the biggest AI-driven hacking incident, but because it illustrates the potential scale of AI-assisted vulnerability discovery and exploitation.
4. There Is a Pattern Here. And It’s Getting Worse.
Let’s connect these three stories.
- Mythos: AI can independently discover previously unknown vulnerabilities in mature, widely deployed software.
- Hugging Face: AI agents can chain vulnerabilities, bypass security boundaries, discover credentials, collaborate, and compromise real infrastructure.
- PHP2Shell: AI-assisted research can accelerate the discovery and development of exploits affecting widely deployed internet infrastructure.
These are different manifestations of the same underlying shift.
AI is changing the speed, scale, and accessibility of cybersecurity operations.
For years, cybersecurity was largely an arms race between human attackers and human defenders, supported by increasingly sophisticated tools. Attackers needed expertise, time, and resources. Defenders had a fighting chance because sophisticated attacks were expensive to develop and execute.
That equation is changing.
An attacker can now use an AI agent to:
- Scan codebases for vulnerabilities.
- Analyze infrastructure and identify exposed services.
- Discover misconfigurations and exposed credentials.
- Generate and test exploit variations.
- Chain multiple weaknesses into a complete attack.
- Automate reconnaissance and follow-up actions.
And they can do it continuously.
But there’s another side to this story.
Defenders can use exactly the same capabilities to identify vulnerabilities, investigate suspicious activity, prioritize remediation, analyze endpoints, and respond to incidents. The problem is that most businesses aren’t equipped to take advantage of this.
They still operate with fragmented security tools, periodic vulnerability scans, manual patching processes, and alerts that nobody has time to investigate. A security team might discover that a server is vulnerable today, but it could take weeks before somebody actually patches it. In an AI-powered attack environment, weeks might as well be forever.
5. What This Means for SMBs
This is where I think the conversation needs to change. Most discussions about AI and cybersecurity focus on large enterprises, frontier AI labs, or sophisticated nation-state attackers.
But small and medium-sized businesses are arguably among the most exposed.
Why?
Because they run the same internet-facing software as large enterprises, but without the same security budgets, infrastructure teams, and dedicated security researchers.
A 50-person accounting firm can run vulnerable PHP infrastructure. A 100-person manufacturing company can have exposed credentials, outdated endpoints, and poorly configured cloud services. A small law firm can have critical customer information sitting behind systems that haven’t been patched in months.
Attackers don’t care whether you’re a Fortune 500 company. They care whether you’re vulnerable.
And AI makes finding vulnerable organizations cheaper. The old approach of buying a firewall, installing antivirus, running an occasional vulnerability scan, and hoping someone patches things on time simply isn’t enough.
Security needs to become continuous.
Not a quarterly project. Not an annual compliance exercise. Not a collection of dashboards that somebody checks when they have time. Continuous visibility, continuous monitoring, and continuous remediation.
6. This Is Exactly Why We Built EspressoLabs
This is also why we’re building EspressoLabs. We believe the traditional MSP and cybersecurity model is fundamentally broken. Businesses are expected to assemble a collection of disconnected products: endpoint protection, vulnerability scanners, patch management, identity management, monitoring, compliance software, and a bunch of consultants to make everything work together. And then somebody is expected to watch all those dashboards. That model wasn’t particularly efficient before AI. It’s going to be completely inadequate in the world we’re entering.
At EspressoLabs, we’re building an AI-native IT, security, and compliance platform designed specifically for SMBs.
The idea is simple: instead of giving businesses another collection of tools, give them an intelligent security and IT operation that continuously understands their environment and acts on it.
Here’s what that means in practice.
1. Know what you have.
You can’t protect what you can’t see.
Our platform focuses on maintaining visibility across endpoints, devices, software, infrastructure, and the security posture of the organization. You need to know what’s running, what’s exposed, what’s outdated, and where your biggest risks are.
2. Find vulnerabilities before attackers do.
AI changes how vulnerability management should work.
Instead of relying exclusively on periodic scans and static vulnerability databases, AI can help correlate software inventory, vulnerability intelligence, configuration weaknesses, and actual exposure.
If PHP2Shell affects your environment, you shouldn’t need to wait for someone to manually review a report.
You need to know which systems are affected, how exposed they are, what needs to be updated, and whether remediation has actually happened.
3. Turn alerts into action.
One of the biggest problems in cybersecurity isn’t a lack of alerts. It’s too many alerts and too little action. AI can help investigate events, correlate signals, distinguish meaningful threats from noise, and prioritize what actually needs attention.
The goal is to move from alert management to actual security operations.
4. Make patching continuous.
The Hugging Face and PHP2Shell stories demonstrate why patch management is no longer an administrative IT task.
It’s a security-critical operation. When a critical vulnerability appears, organizations need to identify affected assets, prioritize remediation, deploy patches safely, and verify the outcome. A 30-day patch SLA might have been acceptable in a different world. It is increasingly difficult to justify when AI can help attackers move much faster.
5. Make compliance part of the operation.
SOC 2, CMMC, and HIPAA shouldn’t be separate paperwork projects.
The same visibility, access controls, monitoring, patch management, and evidence collection that improve security also support compliance. By using AI to automate operational work and evidence collection, we can make compliance faster and more affordable without treating it as a once-a-year exercise. And importantly, AI doesn’t eliminate the need for experienced security professionals, appropriate approvals, or human accountability. It makes those people dramatically more effective.
That’s the model we’re building at EspressoLabs: AI automation combined with real security operations, not another chatbot pretending to be a security team.
7. The Next Cybersecurity Arms Race Is Already Here
I keep coming back to one thought.
The Mythos announcement, the Hugging Face incident, and PHP2Shell aren’t three unrelated security stories.
They’re early signals of a fundamental change in how software is attacked and defended.
AI is becoming capable of discovering vulnerabilities that humans missed for decades. It’s becoming capable of navigating complex infrastructure and chaining exploits together. And it’s dramatically lowering the cost of sophisticated security research. This creates enormous opportunities for defenders, but it also removes many of the barriers that previously protected organizations from sophisticated attacks.
The question for business leaders is no longer whether AI will change cybersecurity.
It already has.
The question is whether your security operation will evolve fast enough to keep up.
At EspressoLabs, we’re betting that the answer is to build security from the ground up around AI: continuous, automated, intelligent, and accessible to businesses that can’t afford a traditional enterprise security team.
Because in this new world, the advantage won’t necessarily belong to the company with the biggest security budget.
It will belong to the company that can identify a problem, understand it, and fix it before someone else finds it.
Welcome to the AI cybersecurity era. It’s going to be an interesting ride.
Want to know how prepared your business is for this new reality?
Visit EspressoLabs.com and see how we’re building the AI-native IT and security platform for the next generation of businesses.
Sources
- Anthropic: Assessing Claude Mythos Preview’s cybersecurity capabilities.
- OpenAI’s incident report
- Hugging Face’s technical reconstruction.
- Bitdefender’s PHP2Shell technical advisory
Discover more from Ido Green
Subscribe to get the latest posts sent to your email.