AI security agents: from buzzword to strategic advantage
AI security agents are specialised automated systems that coordinate multiple machine-driven tasks—such as threat modeling, entry point analysis, and exploit hypothesis testing—to accelerate AI vulnerability detection and code vulnerability discovery across modern software and infrastructure, while keeping human experts in the loop to judge impact and remediation priorities. Agentic tools are no longer a lab curiosity; they are starting to set the pace of enterprise defense. The most striking evidence is Google’s Agentic Vulnerability Discovery Harness (AVDH), an internal chain of AI agents that hunted corporate source code and found over 100 verified, high‑severity flaws in two days during an investigation into stolen repositories. That throughput makes traditional manual security audits and point‑and‑shoot automated security scanning look slow and incomplete. When a tool scans tens of millions of lines of code and produces tens of thousands of findings over ten months, it does more than “assist” humans—it forces CISOs and engineering leaders to reconsider what “reasonable” security coverage should look like.
Inside AVDH: why agent chains beat legacy scanners
The power of AVDH is not magic model quality; it is workflow design. The system runs as a sequence of specialised security agent tools, each handing its output to the next. One agent maps the codebase and produces a threat model, deciding what kind of software it is and which directories to skip. Another scans every file for entry points—web routes, listeners, anywhere user input enters the system. Context enrichment agents then gather scattered pieces of logic—permission checks, sanitizers, and control paths—that a human reviewer would otherwise chase down manually. Hypothesis generation agents look explicitly for access‑control issues like missing authorization or privilege escalation, and for dangerous data flows that lead to SQL injection, cross‑site scripting, command injection, and path traversal. Finally, high‑temperature validation agents argue with each other over each hypothesis before a synthesis agent sorts them into confirmed or rejected. This is automated security scanning, but with reasoning baked into the pipeline. Instead of flagging every suspicious pattern, agents challenge one another and check their conclusions against rules organised by language, framework, and vulnerability type, written by human consultants. That design directly tackles the noisy‑scanner problem that has haunted security teams for years.
From codebases to satellites: AI in real-world incident response
If AVDH proves that agentic AI can tear through code, Atalanta’s Argo shows that AI‑assisted cybersecurity can matter when infrastructure is under active attack. On the day Russia invaded Ukraine in 2022, it disabled thousands of satellite modems across Ukraine and other European countries to disrupt communications. That attack exposed how fragile critical systems can be when a single class of vulnerability is exploited at scale. Argo includes technology that was instrumental in securing Viasat’s satellite communications network after that hacking incident. It is designed in response to continued Russian aggression and newer Iranian threats to cyber infrastructure, aiming to close the kinds of gaps that left satellite links exposed. The subtext is clear: security agent tools are moving from developer laptops into the heart of cyber‑physical systems. Together, AVDH and Argo signal a shift. AI vulnerability detection is not only about finding bugs in web extensions and open‑source projects that later receive CVE assignments; it is also about stabilising communication lifelines when nation‑state actors try to take them down.
Faster detection vs. slower patching: a new tension for enterprises
Enterprise leaders should welcome rapid code vulnerability discovery—but also fear what it exposes about their ability to respond. When a tool uncovers dozens of assignable flaws in widely used extensions and open‑source projects, resulting in 12 assigned CVEs with another dozen in active disclosure, it compresses the vulnerability lifecycle. The discovery phase shrinks from weeks to hours; patch development, testing, and deployment often do not. This creates three uncomfortable questions. First, responsible disclosure: as agentic systems find more issues faster, coordination with vendors and maintainers becomes more complex, especially when findings touch shared dependencies. Second, tooling accessibility: if only a handful of large providers can run these advanced security agent tools, the playing field tilts; attackers will eventually gain similar capabilities, and defenders without AI will fall behind. Third, patch bottlenecks: manual source code review cannot keep pace with AI, and traditional scanning engines consistently miss the broad spectrum of vulnerabilities in modern software. But no AI can yet replace disciplined patch pipelines, change management, and regression testing. Detection without timely mitigation turns into an ever‑growing backlog of known, exploitable weaknesses.
What this means for security posture and developer workflows
Agentic AI will not make security teams obsolete; it will redefine their work. Mandiant’s researchers explicitly encourage defenders who build similar vulnerability discovery harnesses to manually validate findings. In practice, that means humans stay responsible for exploit reproduction, impact assessment, and remediation planning, while agents handle routine pattern hunting and context assembly. For security posture, the implication is blunt: enterprises that do not embed AI vulnerability detection into their code pipelines will be defending at last decade’s speed. Securing the software development pipeline has emerged as a defining challenge in modern enterprise defense, and the success of AVDH shows that defenders can reclaim some advantage against adversarial AI by embedding frontier models inside expert‑defined workflows. Developer workflows will also change. Instead of treating automated security scanning as a noisy checkbox, engineering teams can expect agent chains that feed precise, validated issues into backlogs. That will force harder prioritisation choices but should reduce time wasted on false alarms. The real test is cultural: whether organisations accept that shipping software now includes partnering with AI security agents as routinely as they partner with CI/CD systems. The conclusion is unavoidable: AI security agents are already finding what humans miss, at speeds that rewrite expectations. Enterprises that respond by pairing these tools with disciplined human review and robust patch pipelines will set a new norm for secure software. Those that ignore them will discover, the hard way, that attackers do not wait for manual audits.
