COMMENTARY: Whenever I ask customers what's keeping them up at night, I usually hear some version of "AI-powered attacks."It’s a reasonable concern but pigeonholing agentic AI as a completely new category of threat misses what's actually changed. AI hasn’t replaced traditional attack techniques – at least not yet.It's made them faster and cheaper to run.[SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Read more Perspectives here.] Run autonomous security assessments across enough environments and a pattern emerges. Most findings aren't novel. They're the same issues security teams have been wrestling with for years: SQL injection, misconfigurations, cross-site scripting, exposures that end in remote code execution.Nothing on that list would surprise anyone who's read the OWASP Top 10. Verizon's DBIR shows year-after-year that attackers succeed by exploiting known vulnerabilities, stolen credentials, and configuration mistakes far more often than through anything exotic. AI shortens the timeline. It lets attackers identify and chain those weaknesses in minutes instead of days. Models may appear to surface more "advanced" attacks, but because they can test more permutations at each step than a human ever could. The underlying techniques are still traditional ones, recombined.Today, the economics have changed. Enumerating assets, identifying exposed services, testing configurations, correlating small findings into an attack path — all of it used to take hours of manual work, and increasingly it doesn't. That lowers the barrier for less sophisticated attackers and lets experienced ones operate at a scale that wasn't practical before. Same attack surface, adversaries moving through it much faster.It's better news than it sounds. Rather than contending with two problems, a legacy security problem and an AI security problem, we have one problem that's accelerating. Every unresolved weakness becomes more valuable when attackers can automate reconnaissance, validation, and exploitation. Organizations with mature patch management, disciplined identity controls, and a clear picture of their external attack surface are more resilient because there's less for automated attackers to work with.That said, organizations are exposing infrastructure built specifically for AI workloads. MCP servers are left open without authentication, LLM API keys are left sitting in repositories, and config files and agent identities grant far broader permissions than they need. These risks grow alongside adoption.But they’re familiar problems – missing authentication, exposed credentials, and excessive privileges – in a different context. Too many teams only start asking who owns an agent, what it can access, how its permissions get reviewed, and who's accountable when it makes the wrong call after something has already gone wrong.The teams getting ahead of this are having those conversations before deployment rather than during incident response, and folding AI into existing security and identity governance instead of standing up a separate stack with separate rules.However, AI governance issues are challenging because most companies have yet to work through answers to important questions: Who owns an agent's identity when it misbehaves? Who's liable when an autonomous tool acts on a wrong conclusion? How do we prove our AI's reasoning to an auditor six months later? These questions require time, attention, and cross-functional agreement between security, legal, and the business Most organizations are just starting these conversationAnd we shouldn’t delegate some of these decisions to an agent, regardless of how capable the tooling gets. Humans should remain the ones authorizing scope, irreversible production changes, and final calls on disclosure and materiality. Let agents handle the toil.Still, before buying a new AI security platform, tighten the basics:The AI boogeyman makes a better headline than it does a threat model. A renewed focus on security hygiene may not fill a keynote slot at Black Hat or DEFCON, but it's still where we get the return, because AI-powered attacks depend almost entirely on weaknesses that have been sitting around for years.Klaas Meinke, Head of AI, HadrianSC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Each contribution has a goal of bringing a unique voice to important cybersecurity topics. Content strives to be of the highest quality, objective and non-commercial.
- Improve patching discipline.
- Maintain an accurate asset inventory.
- Review privileged access.
- Rotate credentials.
- Close unnecessary ports.
- Lock down MCP servers.
- Right-size agent permissions.
- Test the external attack surface continuously instead of assuming last quarter's assessment still reflects today's environment.