AI is changing who performs security work, how decisions get made, and what competence looks like. Six connected forces explain why workforce readiness now means developing your people and appraising the agents working beside them.
Imagine watching an aircraft complete a flawless flight on autopilot. It follows the route, maintains altitude, and arrives exactly as expected.
Would that flight alone convince you to trust the system—and the crew supervising it—in a storm?
Putting possible fear of flying aside, you would probably want to know more. Which conditions can the system handle? Can the pilots recognize when it is getting something wrong? And have they practiced taking over when it reaches its limits?
Security leaders face a similar decision as they delegate more work to AI agents. Investigations finish sooner. Reports arrive faster. A promising agent creates a compelling case to expand.
But every expansion also changes what you are trusting the agent to do, and what the person supervising it must be capable of judging. A successful run gives you evidence. It does not establish how either will perform when the conditions change.
Six forces are bringing that question into focus: as AI takes on more security work, how do you know that both the agent and the person directing it are ready for the responsibility?
Finding vulnerabilities has traditionally depended heavily on specialist time. AI agents can now contribute to that work.
In November 2024, Google’s Project Zero and DeepMind reported that their Big Sleep agent discovered an exploitable vulnerability in SQLite. The team described the research as experimental, but the finding demonstrated a concrete capability: an agent identifying a real flaw in widely used software.
That is valuable for defenders. It also challenges the assumption that familiar, extensively reviewed software has little left to reveal.
The leadership implication is practical: discovering more flaws only improves security if people can investigate them, judge their significance, and act. Discovery capacity and response capability need to develop together. Otherwise, the scales can easily tip in the other direction.
Finding a weakness and acting on it are different capabilities. Agents are demonstrating both.
Anthropic reported disrupting a September 2025 espionage campaign in which attackers used Claude Code across reconnaissance, exploitation, and data handling, with limited human intervention. Its account also documented mistakes, including fabricated credentials and incorrect claims about stolen information.
This vendor-reported case does not establish that every adversary is autonomous. It does show why planning exclusively around human-paced execution is insufficient.
For defenders, the question becomes: where does our response depend on someone noticing, interpreting, and approving each step? Exercises should expose those delays and establish which decisions can be made safely in advance.
As agents perform more tasks, practitioners must decide what to delegate, how to direct it, and whether to accept the result. At least it won’t eat the lunch you left in the fridge with your name on it.
Consider a hypothetical penetration test. An agent proposes a technically valid action against a system outside the agreed scope. The operator who approves it may understand the technique perfectly and still make the wrong decision.
Competence therefore includes technical understanding, scope awareness, verification, and intervention. Knowing when to let a sound action proceed matters too: unnecessary overrides can make delegation ineffective.
Role development should reflect that responsibility. People need opportunities to practise supervising agents, explain their decisions, and demonstrate that they understand the work being performed.
An agent that takes instructions, uses tools, and produces work of variable quality needs an accountable owner and a clear performance standard. Otherwise it won’t take home that end-of-year bonus.
Once agents perform security tasks, your team contains two kinds of worker: human and agentic. That does not transfer accountability away from people. It expands what people must manage.
It also changes how experience develops. If agents absorb routine investigations, junior practitioners may lose some of the repetitions through which they previously learned. That outcome is not inevitable, but avoiding it requires deliberate design.
Ask two workforce questions together: who owns the quality of delegated work, and where will people gain the experience needed to challenge it?
Related reading
Want to explore the missing middle?
Read about the AI silent killers your dashboard will never show you.
Read the articleA policy can specify that a person must review an agent’s findings. It cannot establish whether that person can recognize a convincing error.
NIST’s voluntary AI Risk Management Framework treats measurement as an ongoing activity informed by deployment context. Its supporting guidance addresses evaluating systems in conditions resembling their intended use and tracking risks over time.
For security leaders, the useful distinction is between defining expectations and demonstrating that they hold.
Evidence should identify the task, operating conditions, standard, result, and limitations. A successful exercise supports a bounded conclusion about capability. It does not prove readiness for every incident or establish compliance with every obligation.
An agent can keep working without producing a useful result. Repeated attempts, unnecessary investigation, and human rework all affect the economics of delegation.
Imagine two hypothetical investigations that reach the same correct conclusion. One requires minimal intervention. The other consumes repeated model calls and extensive analyst correction. Counting completed investigations makes them look equivalent.
Basically, some of your workers no longer receive a monthly salary; they run on a meter. Leaders need to connect cost to verified outcomes, including the effort required to check and correct the work.
That makes cost discipline part of competence. An effective operator knows when to continue, change direction, intervene, or stop.
These forces converge on a single management problem: responsibility is moving faster than the evidence used to assign it.
Start with one consequential workflow and apply five disciplines:
Hack The Box brings this discipline into Cyber Workforce Development for Human and Agentic Teams: developing the people who direct AI and appraising the agents working beside them.
For the people directing AI, completing the task is only part of the test. HTB makes their judgment measurable:
Together, these measures help answer a question adoption figures cannot: Who can you trust to direct AI, with what authority, and on what evidence?
The agents themselves need to earn that trust too. With HTB AI Range – Enterprise Edition, organizations can:
Your security workforce now has two kinds of worker. You need evidence that the agent can do the job—and that the person directing it knows when to intervene.
Before handing over more responsibility, ask: Has the agent earned that responsibility? Has its operator demonstrated the judgment to supervise it?
Build a workforce ready for human–AI collaboration
Explore Cyber Workforce Development for Human and Agentic Teams with Hack The Box.
Talk to Hack The Box