If the queue is clear and the dashboard is green, why does something still feel wrong?
In medicine, a silent killer is dangerous precisely because it doesn’t announce itself. It develops without obvious symptoms, while the things you normally use to judge whether someone is healthy continue to look normal.
AI may be creating a similar blind spot inside your organization. Ask a CISO what AI has changed and the answer will often start with the attacker: faster exploits, autonomous scanning, deepfakes, attacks that scale beyond what a human team can match.
All true, but there is another side to the transformation that’s much harder to see. AI is changing how security teams work, what people practice, and which roles develop experience. At the same time, the security team itself is changing. Your people now work alongside agents that can perform tasks, make decisions and produce outcomes. None of those changes trigger an alert, but they can weaken your workforce and place your security at risk.
Imagine receiving a ticket that your AI agent flags as low risk. Do you accept its recommendation without a second thought or review it first?
If you accept it, everything looks normal. The ticket moves through the queue and the SLA is met.
But what if the agent was wrong? The problem isn’t just the incorrect recommendation. It’s that nobody challenged it.
That’s automation bias: trusting an automated recommendation simply because it came from the system, even when it should have been questioned.
Research into automation bias in human-AI decision-making found that as AI agents perform more of the work, people are more likely to rely on their outputs, especially when they trust the system, are under time pressure, or when verification takes effort. That creates a difficult paradox: the faster the workflow becomes, the fewer opportunities people may have to exercise their own judgment.
How to spot it: Put a wrong AI agent decision in front of the operator. Do they catch it?
You probably rely on an AI agent for part of your work every day. If it disappeared tomorrow, could you still do that work yourself with the same speed and confidence?
You'd probably remember the tools and the workflow. The harder question is how much practice you've had recently.
When an agent handles a task day after day, the person using it gets fewer opportunities to perform the underlying work themselves. Over time, those repetitions disappear from the job and the skill becomes harder to maintain.
The impact may take months or years to become visible. A team can become faster in the short term while gradually losing opportunities to practise the skills that allow people to understand, verify, and take over the work when necessary.
How to spot it: Take the agent away. Can the analyst still do the work unaided, and can they still judge whether the agent's answer is right for this environment?
You’re using AI to take repetitive, entry-level work off your team’s plate. The immediate benefits are easy to see: fewer repetitive tasks, faster resolution, and more time for higher-value work.
Then consider what used to happen during that work. A junior analyst worked through alerts and gradually learned to recognize patterns. A new engineer troubleshot simpler issues before moving on to harder ones.
When an agent handles those tasks instead, the development path changes.
Where will your next experienced practitioners come from?
AI can help people perform beyond their current skill level. That can be valuable for the person and the organization. The same process can also remove some of the experiences that helped people develop into senior practitioners in the first place.
HTB’S Neurogrid report shows the other side of the equation. AI-augmented teams had a 3.2x higher average solve rate than human-only teams, but that advantage narrowed as expertise increased, from 3.2x across the full cohort to 1.69x among the top 5%.
A junior practitioner can therefore become more productive with AI while getting fewer opportunities to build the experience that eventually makes them senior. The work still gets done, but part of the development process has disappeared.
That’s the Missing Middle: the gap between the work organizations automate today and the experienced people they will need tomorrow.
How to spot it: Ask where your senior analysts in 2031 are coming from.
AI is the worker nobody puts on the organization chart. It passed its evaluations, performed well during procurement, and handled the job it was brought to do.
But the environment around it doesn’t stay still. The model, prompts, or surrounding systems may have changed.
A human employee has regular opportunities for feedback, assessment, and retraining. An agent doesn't automatically get any of those things. Without ongoing assessment, there is little to tell you that its capability has moved. An agent that handled 95% of tier-one triage at go-live might be handling 80% today, and you won't know unless you reassess it.
Monitoring can tell you whether the agent is running. It can't tell you whether it is still doing the job to the standard it was originally approved against. That’s silent agentic drift.
How to spot it: Rerun the original appraisal against the current AI agent.
These failures point to the same underlying shift: AI has changed the security workforce itself. So measuring whether the work gets done is no longer enough. You need to know whether the people doing it are capable, whether the AI they rely on is being used effectively, and whether the agents performing the work can actually do their jobs to the standard you expect.
Hack The Box is the only platform that develops your people, appraises your agents, and helps you measure both.
We do this across three connected areas:
Security for AI: Develop the practical skills needed to understand, attack, and defend AI systems, so your teams can secure the AI your organization deploys and defend against AI-powered threats.
AI for Security: Help security practitioners use AI effectively in their work while maintaining the judgment and technical skills needed to direct it, challenge its decisions, and intervene when it gets things wrong.
Agentic worker competence: Evaluate whether both people and AI agents can actually perform cybersecurity work to the standard required, rather than relying on productivity metrics or a one-time assessment.
This means giving practitioners hands-on environments where they can build and demonstrate real capability, while giving organizations a way to assess the performance of the agents they increasingly rely on.
Explore how HTB can help develop your people, appraise your agents, and measure both.