
October 15, 20262 min readXY.AI Labs Team
Introducing XY MCP. Connect Claude Code or any MCP client to XY agents, Knowledge Bases and integrations.
Discover9 min read
XY.AI Labs Security Team

Insights from our team on why the rise of autonomous AI agents demands a fundamental rethink of cybersecurity and how we're building infrastructure designed for this new reality.
Our Head of Security recalls a memorable lesson from a college combinatorics course called Introduction to Counting.
During a discussion about securing computers, Professor Barefoot offered an unconventional solution:
"Anything less," the professor explained, "is a compromise."
The lesson wasn't about physical security. It was about understanding that security is not binary. It is a problem of probabilities, attack surfaces, and trade-offs.
In combinatorics, risk can be understood through the number of possible outcomes. The objective isn't to eliminate every conceivable attack path—that would require burying our computers in concrete. It's to reduce the number of viable paths, increase the difficulty of exploitation, and minimize the potential payoff.
For decades, cybersecurity strategies have been built around these principles. But agentic AI is fundamentally changing the equation.
Recent security incidents and disclosures illustrate how vulnerabilities can emerge across nearly every layer of the technology stack.
These are different attack vectors, but they share a common implication: defending individual systems is no longer enough when attackers can exploit the connections between them.
Agentic AI makes this challenge considerably more urgent.
Historically, cyber threats often fell into two broad categories: relatively unsophisticated automated attacks exploiting known vulnerabilities, and highly skilled, well-funded teams capable of chaining together complex exploits.
Autonomous AI agents are narrowing the gap between those categories. An attacker no longer necessarily needs to manually investigate every weakness or develop sophisticated exploits from scratch.
AI agents can increasingly assist with analyzing newly released patches, exploring network configurations, identifying vulnerable dependencies, combining minor weaknesses, and generating personalized social-engineering campaigns at scale.
The result is a fundamental shift in the economics of cyberattacks. What once required significant expertise, time, and coordination is becoming faster, cheaper, and increasingly automated.
For a company like XY.AI Labs, operating in highly regulated healthcare environments, this changes how security must be designed. The challenge is no longer simply defending against static scripts or occasional sophisticated adversaries.
It is preparing for persistent, moderately sophisticated autonomous operators capable of working at machine speed.
Consider an increasingly plausible scenario.
A software developer receives what appears to be a legitimate opportunity to collaborate on a technical project. The invitation includes a GitHub repository containing a coding assignment.
The repository looks credible. The instructions are familiar. The developer clones it and runs a standard command:
npm installUnknown to the developer, the repository contains a malicious installation script. The moment the command executes, the script begins collecting information from the workstation.
It searches for environment variables, API credentials, locally stored secrets, cryptocurrency wallets, and password stores. It then attempts to transmit that information to an attacker-controlled server.
This type of attack is particularly dangerous because it exploits a routine development workflow rather than a sophisticated vulnerability in the operating system.
Now imagine that several protective conditions happen to limit the damage:
In this hypothetical scenario, the immediate damage might be limited. But the outcome would depend partly on luck.
The attacker-controlled server could have been available. Valuable credentials could have been exposed. An accessible API token could have enabled lateral movement into other systems.
Luck is not a security architecture.
A security strategy cannot depend on developers never executing malicious dependencies, attackers' infrastructure being unavailable, or defenders consistently patching vulnerabilities faster than automated tools can exploit them.
Instead, the more useful question is:
How can organizations systematically reduce the probability and impact of compromise without making their developers' computers unusable?
At XY.AI Labs, our Head of Security's approach centers on two architectural priorities.
Many organizations respond to rapidly emerging vulnerabilities by accelerating their patching processes. This remains essential security hygiene.
But relying primarily on patching speed creates an increasingly difficult race. AI-assisted attackers can analyze disclosed vulnerabilities and develop exploitation techniques rapidly, while responsible engineering organizations still need time for testing, validation, and deployment.
Our approach is to complement patching with something more fundamental:
Reduce the production attack surface so dramatically that an attacker has fewer tools and opportunities available after gaining access.
At XY.AI Labs, our production architecture emphasizes minimal containers built using Docker's FROM scratch approach.
These containers are designed to contain:
This is a deliberate departure from traditional application containers. A conventional container may include a shell, package managers, networking utilities, scripting runtimes, and other tools.
If an attacker gains code execution, those tools can become part of the attack. They may attempt to launch /bin/sh, download additional payloads using curl or wget, execute scripts, inspect their environment, or establish persistence.
Security professionals refer to this technique as living off the land: using tools already available inside the compromised environment.
In a minimal scratch container, many of those tools simply do not exist.
There is no shell. There is no package manager. There is no curl or wget. And a read-only root filesystem prevents straightforward attempts to write persistent payloads into the container.
This does not eliminate every possible attack. Vulnerable application code may still expose data, make unauthorized network requests, or access resources permitted by its runtime identity. But it substantially reduces the attacker's available options.
The objective is not to make exploitation mathematically impossible. It is to make successful exploitation significantly less useful.
This approach creates operational trade-offs.
Developers cannot simply open a shell inside a production container to investigate a live problem. Instead, engineering teams must rely on structured telemetry, logging, observability, and local reproduction.
That requires discipline and investment. But it also creates an important benefit: security policies become architectural properties rather than guidelines that depend on consistent human behavior.
Many organizations already discourage direct developer access to production environments. Minimal, immutable containers help enforce that principle technically.
Hardening production changes where attackers are likely to focus. When production environments become more difficult to exploit, developer workstations remain attractive targets.
A developer's laptop is almost the opposite of a minimal production container. It needs compilers, package managers, development tools, API clients, browser sessions, AI coding assistants, and access to multiple services.
Those capabilities are essential for productivity. Eliminating them would recreate the professor's concrete-filled security solution: technically restrictive, but practically unusable.
Our Head of Security advocates a different principle:
Assume that a developer workstation may eventually be compromised, and design credentials so their exposure has a limited lifespan and blast radius.
Let's Encrypt and Certbot helped transform how organizations manage TLS certificates. Instead of relying on long-lived certificates and manual renewal processes, organizations increasingly adopted automated issuance and renewal.
The same principle can be applied to infrastructure secrets, API keys, and access tokens.
At XY.AI Labs, this philosophy informs an ongoing effort to make credential rotation routine, automated, and operationally reliable. Two capabilities are central to that approach.
Our approach is to treat infrastructure credentials, API keys, and access tokens as short-lived assets that require regular, automated rotation.
If credentials are silently compromised, limiting their validity reduces the window of opportunity for bad actors. Equally important, automated rotation builds operational resilience: services must be designed to accept refreshed credentials seamlessly, without downtime or manual intervention.
The goal is to make credential rotation a routine, reliable, and continuously validated security practice rather than an exceptional response to an incident.
The second capability builds on the first. If every secret is managed through a reliable automation framework, emergency response becomes dramatically faster.
Imagine a developer accidentally executes a malicious package.
Instead of spending hours determining which credentials may have been exposed before beginning remediation, a security team should be able to trigger coordinated credential revocation and regeneration across affected systems.
The objective is a one-button emergency reset that minimizes the time between detecting a compromise and invalidating exposed credentials.
That capability requires careful engineering, dependency management, and testing to avoid interrupting critical services. But it changes the economics of incident response.
Attackers benefit when stolen credentials remain valid for long periods. Defenders benefit when those credentials can be invalidated quickly and systematically.
This architecture is still evolving.
Some third-party platforms offer mature APIs for credential management. Others make automated rotation surprisingly difficult, requiring manual browser interactions or administrative workflows.
Even with modern AI agents capable of automating browser-based tasks, managing these processes reliably can remain challenging. In some cases, the inability to automate secure credential lifecycle management may influence vendor selection.
We believe credential automation should become a standard expectation for enterprise software infrastructure.
For developers and partners building on XY.AI Labs, our developer platform provides credential management capabilities at xy.ai/developers/credentials. These capabilities can support the development of automated credential-management workflows.
The professor's lesson remains relevant.
There is no such thing as a completely secure computer that is also useful. Every system involves trade-offs.
Agentic AI is making those trade-offs more consequential because attackers can increasingly automate activities that previously required significant time and expertise.
At XY.AI Labs, our security philosophy is therefore not built around the assumption that every attack can be prevented. It is built around reducing the attack surface, limiting privileges, shortening credential lifetimes, and making recovery faster and more predictable.
None of these controls provides perfect protection independently. Together, they create a more resilient architecture.
As agentic AI becomes more capable, cybersecurity must evolve from a collection of defensive policies into a system designed to remain resilient even when individual protections fail.
The goal is not to build an impenetrable fortress. It is to ensure that when an attacker finds a way inside, there is as little as possible to exploit, as little as possible to steal, and as little time as possible to act.
Check our XY.AI Labs Trust Center. Our commitment to data privacy and security is embedded in every part of our business. Use the XY.AI Labs Trust Center to learn about our security posture and request access to our security documentation.
Book a DemoSee how AI Agents can transform your operationsBlog

October 15, 20262 min readXY.AI Labs Team

October 12, 20263 min readXY.AI Labs Team

October 12, 20264 min readXY.AI Labs Team

September 25, 20263 min readXY.AI Labs Team

September 21, 20263 min readXY.AI Labs Team

September 21, 20265 min readXY.AI Labs Team

August 28, 20264 min readXY.AI Labs Team
August 14, 20266 min readScott Cressman

August 3, 20268 min readSam De Brouwer

July 9, 20265 min readXY.AI Labs Team

June 25, 20263 min readXY.AI Labs Team

June 18, 20268 min readSam De Brouwer

May 27, 20264 min readXY.AI Labs Team

May 13, 20266 min readSam De Brouwer

May 6, 20265 min readXY.AI Labs Team

April 30, 20264 min readXY.AI Labs Team

April 29, 202610 min readXY.AI Labs Team

April 15, 20264 min readXY.AI Labs Team

March 11, 20265 min readSam De Brouwer

February 21, 20264 min readXY.AI Labs Team

February 5, 20265 min readXY.AI Labs Team

January 29, 20265 min readSam De Brouwer

December 19, 20257 min readSam De Brouwer

November 24, 20258 min readXY.AI Labs Team

November 13, 202510 min readSam De Brouwer

November 10, 20256 min readSam De Brouwer

October 13, 20258 min readSam De Brouwer

September 12, 20255 min readScott Cressman
August 28, 20253 min readSam De Brouwer
August 24, 202510 min readXY.AI Labs Team

August 23, 202510 min readXY.AI Labs Team
August 22, 20252 min readXY.AI Labs Team
August 14, 20254 min readSam De Brouwer
July 31, 20256 min readSam De Brouwer
July 10, 20254 min readSam De Brouwer
June 10, 20255 min readXY.AI Labs Team
May 6, 20258 min readSam De Brouwer
April 17, 20258 min readScott Cressman

March 24, 20259 min readSam De Brouwer
Get started now