How should security teams respond to agentic AI?
AI is changing both sides of security.
Attackers can find and exploit vulnerabilities faster, and defenders have new ways to fight back. Security teams can respond by tightening vulnerability ownership, treating AI agents as privileged identities under zero trust, using internal data to power defensive AI, and automating repeatable work while keeping experts on judgement calls.
In Episode 1 of The Stack, Montash's Alex Reed spoke with Mattia Bradascio, Founder & CEO of Autopoiesis, about how AI is changing cyberattacks and what security teams need to change in response.
They covered:
- How AI-enabled cyberattacks have moved beyond generated phishing towards autonomous vulnerability discovery and exploitation
- How zero-trust principles, identity, sandboxing and runtime monitoring apply to autonomous agents
- How security teams can use private data to give defensive AI an advantage over attackers
- Which security work AI can automate, and which still needs people
- Emerging AI security research, including model verification, recursive self-improvement and hardware-level monitoring
Here are four key takeaways from Episode 1.
1. AI security starts with the foundations
Before adding autonomous agents to your security stack, Mattia recommends testing the basics:
“If some agent... was able to find a critical vulnerability in one of your systems today, would that organisation actually know who owns that component in that system? And how quickly do you think you could get that fixed, released into production?”
As AI increases the speed and volume of security work, organisations need clear ownership and a reliable route from discovery to remediation.
2. Treat AI agents like privileged users
Agentic systems create a different security challenge because they can act across your environment, not just advise.
Mattia compared a capable AI agent to "a very dangerous human": something that can act independently, at speed, across your systems.
His starting point is zero trust. Give each agent its own identity, control what it can access and apply security policies to its actions. Sandboxing can isolate the agent from critical systems, while runtime monitoring helps teams identify dangerous behaviour and intervene.
Without those controls, a capable agent with excessive access creates risk at machine speed.
3. Use the information attackers don't have
External attackers see your systems from the outside. Your security team has access to source code, architecture, known vulnerabilities and internal risk data.
Mattia sees that private information as an advantage:
“I think the trick is leveraging that and basically building upon your asymmetric information advantage.”
One application is continuous purple teaming. Security agents could repeatedly search for vulnerabilities and speed up remediation using information organisations would never expose publicly.
Attackers don’t have exclusive access to capable AI. Security teams can use the same technology, with far more context about the systems they’re defending.
4. Automate the repeatable, keep experts on the judgement calls
AI agents can already accelerate coding, source-code analysis and established security workflows.
Mattia was clear about the limit:
“I don't think we're at the point yet where human experts are entirely replaced.”
Current agents still struggle with security work that requires stealth, judgement and novel approaches. Mattia pointed to repetitive and well-defined tasks as stronger candidates for automation, freeing specialists for the work only they can do.
That changes how security teams use their people. It does not remove the need for them. For security leaders, that means rethinking which skills to hire for as agents take on more of the repeatable work.
What else does the full episode cover?
The conversation goes further than the four themes above.
Mattia and Alex discuss how AI-enabled attacks have developed from relatively simple uses of language models into increasingly autonomous security systems.
They also explore what an internal agentic security platform could look like, including how organisations bring vulnerability data, threat intelligence and other sensitive security information into one controlled environment.
Later in the episode, Mattia moves into the research shaping what comes next: verifying AI models running inside sovereign infrastructure, controlling recursive self-improvement, hardware-level monitoring and the standards that could govern increasingly capable AI systems.
For CISOs, cyber leaders and technical teams, the full episode goes deeper into what it takes to use AI inside a security environment without introducing more risk than it removes.
Watch The Stack: Episode 1
Catch the full conversation with Mattia Bradascio for more on agentic security, autonomous defence and where AI security is heading next.
Keep up with The Stack
The Stack brings technical leaders and practitioners together to discuss the technologies changing how organisations build, secure and scale.
Follow Montash on LinkedIn to hear when the next episode is announced.
