For ten years, AI in security was a smart filter. It ranked alerts. It flagged anomalies. It wrote summaries. A human still did everything that mattered.
Agentic AI is different. It plans. It acts on live systems. It looks at the result and decides the next step. No human at every turn.
Vulnerability discovery has become a compute problem
In 2025 and 2026, multiple autonomous AI pentesters reached #1 on HackerOne’s US leaderboard, ahead of thousands of human researchers. Autonomous agents are now competing with leading human researchers at finding bugs in real production code.
For twenty years, finding a new bug was slow, expert work. Only elite humans could do it. Now an agent with a good model does it in hours. Talent was the bottleneck. Compute is cheap.
The first targets are custom web apps and APIs. Nobody outside the company has tested that code. The bugs are still there. And the attacker no longer needs one big flaw. Three small ones, chained, are enough.
Meanwhile, patching is as slow as ever. Open-source maintainers have asked AI labs to slow down disclosures because they cannot keep up. The bottleneck has moved from finding bugs to fixing them.
Security assessment moves from annual to continuous
Most companies test their defenses once a year. That rhythm made sense when attackers needed months to find a bug. Now they need hours.
Coverage is thin. A typical programme tests about a fifth of its applications. Scanners cover the rest and drown teams in false positives. Scanners do not find business logic flaws at all, and those are the ones that hurt.
Agentic AI gives defenders the same leverage it gives attackers. Agents find the assets. Agents test them every week, not every year. Agents chain findings into real attack paths. And agents prove each one with a working exploit, not a guess.
Three things have to go up together. Scope: test everything, not a sample. Depth: reach the business logic and the multi-step workflows. Speed: test every release.
Reasoning is not execution. A model can propose an attack. It cannot, on its own, guarantee the action was safe, in scope, repeatable and logged. Hypotheses are not vulnerabilities. The hard engineering sits around the model: scope control, rate limits, no writes by default, a full audit trail. Non-deterministic AI needs deterministic guardrails.
Beyond testing: operations, remediation, attack and identity
Testing is one area. The same shift is under way in four others.
Defence. SOC analysts spend most of their day on triage. Agents are already doing the first pass: pulling context, closing false alarms, escalating the real ones. The analyst becomes the reviewer.
Fixing. Agents that find a bug and prove it will also write the patch, test it, and open the pull request. The remediation bottleneck is the next one to break.
Attack. Phishing was a numbers game. Agents make it a precision game. One agent can research a target, write a personal message, and hold the conversation.
Identity. Every agent is a new user with credentials and permissions. Machine identities already outnumber humans 109 to 1, and companies now have to govern machine accounts that act on their own. Agents are also a new target. Prompt injection is the SQL injection of this decade.
What the next five years look like
Agentic AI will automate most of the routine work in security: alert triage, vulnerability validation, patch generation, access reviews, compliance evidence. Gartner expects agents to make 15% of day-to-day work decisions by 2028, up from zero in 2024. Security will not be exempt.
The same agents are the next class of risk. Machine identities already outnumber humans 109 to 1, and AI agents are the fastest-growing part. Gartner has made agentic AI oversight its top security trend for 2026. Forrester predicts an agentic AI deployment will cause a public breach this year. The guardrails are not built yet.
The jobs change too. The analyst who read alerts will supervise the agents that read them. The pentester who ran tools will design attacks for agents and judge the results. The scarce skill moves from doing the work to governing it.
Efficiency is the first-order effect. New risks and new roles are the second. Both arrive together. Companies that treat agents as a new workforce, with identities, limits, and accountability, will get the upside. The rest will find out the hard way.
