AI Is Reshaping Red Teaming Through Faster Reconnaissance and Testing

AI is changing offensive security by reducing the time and labor required to prepare, run and document red team exercises. The practical effect is already visible in the figures shared by DayBlink Consulting: work that once took a small team months can, in some cases, produce dozens or hundreds of findings in a single exercise. For security leaders, the challenge is no longer just how to use AI defensively. It’s how to keep testing programs relevant as attackers gain the same speed advantage.

That shift calls for a distinction that often gets lost in cybersecurity discussions. AI red teaming tests language models and agentic systems for weaknesses, including prompt-injection vulnerabilities. Red teaming with AI uses AI as a tool for conventional offensive operations. The second is changing day-to-day security work most directly: it helps operators gather intelligence, adapt tests and handle the volume of reporting that follows an engagement.

Speed changes the scale of an engagement

The reported productivity gap is substantial, though the figures describe particular engagements and shouldn’t be read as universal benchmarks. DayBlink Consulting aggregated results from ten years of work and says Robert Pimentel confirmed that the comparisons reflect his red team experience.

Those comparisons point to a changed operating model. Automation can widen the scope of an exercise and shorten the time between a question and a result. Yet speed alone doesn’t establish whether a finding is valid, useful or safe to act on. Teams still need operators who can choose targets, understand system behavior and check what automated tools produce.

AI now supports the full engagement cycle

AI’s role has moved beyond drafting phishing messages. Both organizations and threat actors are applying it across operational stages, as described in this account of AI use in espionage. The degree of automation varies by task: preparation and documentation can be delegated more readily, while decisions during an active operation still call for experienced human judgment.

• Preparation and reconnaissance: Tasks that could occupy weeks, such as collecting open-source intelligence, mapping an organization, provisioning infrastructure and configuring command-and-control servers, can now be accelerated with agentic AI. Experienced operators still need to guide complex infrastructure choices. The gain is especially practical in setup work: teams can spend less time assembling the operation and more time making decisions about its purpose and scope.
• Enumeration and attack: During an engagement, operators can use AI to write custom audit scripts, investigate unfamiliar technology stacks and adjust their approach as new information emerges. One described safeguard is to have a second model review another model’s output for correctness and risk before execution. That can improve confidence in automated steps, but it doesn’t remove the need for a human to assess whether an action makes sense in context. The attack phase remains the part where operator judgment and experience matter most.
• Reporting and deconfliction: After an operation, AI can turn large volumes of raw data into a useful first draft. The source describes processing nine to ten months of command-and-control logs in under 20 minutes, mapping commands to the MITRE ATT&CK framework, and organizing information for executive readers. A model can save days of drafting, but people must verify the findings, check the framework mappings and confirm that the final account is accurate. Review is part of the work, not an optional finishing step.

Human judgment separates useful automation from noise

AI can handle more basic scanning, enumeration and scaffolding, but those tasks aren’t the whole job. Effective red teaming depends on understanding architecture, identifying where trust boundaries break down and deciding which weaknesses matter. The strongest operators use AI to move faster on those assessments. They don’t hand over the reasoning itself.

That distinction creates a widening skills gap. At one end are practitioners with technical depth and adversarial judgment who can direct AI toward a specific question and assess its output. At the other are operators who paste generated results into a workflow without checking how they were produced. The latter risk generating inaccurate, noisy activity that wastes time and draws attention without advancing the exercise.

Hiring needs to account for this change. Certifications can still provide useful evidence of training, but they can’t show on their own whether someone can reason through unfamiliar systems or evaluate AI output. Teams should also look for curiosity, an understanding of network architecture, hands-on tool building and the ability to question a model’s suggestions. The value lies in pairing technical fundamentals with the judgment to decide when automation is useful.

What security leaders should prioritize

The next 18 months will test whether organizations can adapt their security programs to faster, more automated attacks. Several priorities follow from the operational changes described here:

• Budget for model access and experimentation. Frontier models can be expensive to use. Security teams need enough access to try them, understand their limits and build relevant workflows.
• Make adversarial testing continuous and programmatic. As enterprises add thousands of agentic integrations, occasional manual exercises won’t cover every system or change. Automated testing can help teams assess systems more regularly.
• Account for attacks that combine digital and human methods. As automated defenses make some digital approaches harder, operations may mix technical techniques with social engineering or physical access attempts.
• Hire for active learning and sound judgment. Teams need practitioners who experiment with AI and can direct it responsibly, alongside the technical knowledge required to assess the results. This discussion of cybersecurity skills in the AI era makes a similar case for prioritizing capability.

AI gives offensive teams more capacity, but the largest gains depend on how well people direct and verify the work. Security leaders should treat that combination as an operational requirement: equip teams to test faster, build review into every automated step, and develop the human expertise that turns a flood of output into defensible findings.