AI Is Reshaping Red Team Operations, From Reconnaissance to Reporting

AI is changing the scale and pace of offensive security. The most immediate shift isn’t AI systems being tested for flaws; it’s operators using AI to carry out familiar red team work faster, across more targets, and with less manual effort. For security leaders, that raises a practical question: can their testing and response practices keep pace with operations that used to require far more people and time?

DayBlink Consulting’s figures, gathered across a decade of engagements and described as representative of Robert Pimentel’s red team experience, suggest a sharp change in output. A three or four person traditional team might produce 60 to 70 findings in a year. The article reports that an AI augmented exercise can generate more than 90 findings in a single after hours session, while an engagement on one application may surface over 100 true positive findings, compared with the traditional range of seven to 10.

These numbers are striking, though they should be read as reported field results rather than a universal benchmark. They point to a shift in operating economics: small teams equipped with frontier models can tackle work that previously demanded lengthy manual workflows. That creates pressure on defenders, especially where security reviews still depend on periodic, labor intensive testing.

Two different uses of AI in security testing

Industry discussion often groups distinct activities under the label of AI red teaming. One practice tests large language models and agentic pipelines for vulnerabilities, including prompt injection and unintended behavior. The other uses AI as a tool for traditional offensive security operations. The latter is where the source article sees the most immediate operational change.

That distinction matters for planning. Testing an AI product’s behavior calls for methods suited to model and application risks. Using AI to support an attack simulation affects reconnaissance, infrastructure setup, technical investigation, and reporting. Security teams need to account for both, but the second changes the tempo of work across an entire engagement.

More capacity across the engagement

Before an operation begins, AI can take on parts of reconnaissance and setup that once consumed days or weeks. It can help gather open source intelligence, assemble organizational information, provision infrastructure, and configure command and control systems. Experienced operators still need to make decisions about complex infrastructure, but automation can clear routine work and get an engagement underway sooner.

During an active operation, practitioners can use models to draft custom audit scripts, investigate unfamiliar technology, and adapt their approach as new information appears. The value here is not simply generating code. It’s having a tool that can help an operator move quickly through technical questions while maintaining control over what gets executed. One workflow described in the source has a second model review the first model’s output for correctness and risk before action. That adds a check, but it doesn’t remove the operator’s responsibility to assess the result.

At larger scales, the reported difference is especially pronounced. Traditional scanning may require careful sequencing and established pipelines, and can time out. With AI supported capability building, the article says teams can scan millions of IP addresses in two to three hours, at a cost of roughly $20,000 to $30,000. The article also reports that external application reconnaissance can yield more than 100 findings a month, compared with 150 to 200 valid findings over a traditional year. The figures underline the possibility of greater throughput, while also making quality control central: more output is useful only when teams can confirm what is real and relevant.

Reporting gets faster, but still needs review

Post operation reporting has long involved a heavy manual workload. The article describes feeding nine to 10 months of raw command and control logs into a frontier model to produce a deconfliction report in under 20 minutes. The model can map commands to the MITRE ATT&CK framework, compare activity with threat intelligence, and turn technical records into an executive facing account.

That changes the time required to reach a first draft, from days to minutes in the example provided. It doesn’t make the report ready to deliver. Operators still need to verify the activity, check ATT&CK mappings, and confirm that the account is accurate. The practical gain is time returned from routine documentation and ticket follow up, which can be redirected toward analysis and remediation.

Human judgment sets the ceiling

Automation can help with enumeration, scanning, and the scaffolding around an operation. It can’t substitute for understanding how a system is built, where trust boundaries break down, or which apparent weakness can become a meaningful exploit. The strongest practitioners will pair that foundation with AI tools and use them to move faster on informed hypotheses.

The article draws a sharp distinction between those operators and people who treat model output as ready made work. Practitioners who copy and paste unverified results risk generating noisy activity that defenders can detect and that wastes time. In contrast, operators who understand the underlying systems can use AI to handle repetitive tasks while focusing their own attention on complex exploitation and judgment calls.

That has implications for hiring. Certifications remain relevant, but they can’t be the sole measure of readiness. Teams should also look for hands on curiosity, knowledge of network architecture, critical thinking, and experience building tools with frontier models. The key question is whether a candidate can direct the technology and check its work, not simply operate an interface.

What security leaders should prepare for

The article’s priorities for the next 18 months include setting aside formal budgets for access to frontier models, since teams need room to experiment and build. It also calls for continuous, programmatic adversarial testing as enterprises deploy more agentic integrations. Periodic manual assessments are unlikely to cover thousands of systems and connections at the pace they change.

Leaders should also expect offensive activity to combine automated digital techniques with physical and social engineering. As automated defenses strengthen digital perimeters, operations may mix technical exploits with efforts aimed at people and physical access. That combination makes integrated testing more useful than treating digital and human risks as separate exercises.

The operational takeaway is straightforward: AI expands what a red team can attempt, but it doesn’t make expertise optional. Security programs that pair automation with experienced judgment can test more frequently and produce results sooner. Those that rely on manual processes alone may find their assumptions outdated before the next assessment begins.