Cybersecurity AI: what is actually working now?

Why is cybersecurity AI suddenly worth taking seriously?

Cybersecurity AI has clearly moved into real security work now, even though most companies are still figuring out how far they can safely push it.

The latest SANS AI survey gives us a good starting point. Among 536 cybersecurity and IT practitioners, 78% said their teams were using AI in cybersecurity, up from 50% one year earlier. AI use in red-team work went from 33% to 61% over the same period, while investigation, response and application security also became common areas of deployment.

The more interesting number is 27%. That is the share of respondents who described their AI deployments as mature production systems. Most teams are therefore somewhere between experimentation and partial operational use.

At the same time, 63% said they had encountered significant shortcomings with AI in threat detection and response, compared with 45% one year earlier.

That is why the debate still feels messy. Adoption is moving much faster than trust. Security teams are already getting useful work from AI while also finding, sometimes painfully, where it still breaks.

What does it actually mean for cybersecurity AI to “work”?

Cybersecurity AI is working today when it reliably removes human work, improves a security decision, finds something humans were missing, or makes an incident meaningfully faster to contain.

That definition rules out a lot of impressive-looking demos.

A chatbot explaining malware is convenient. An AI system that correctly sorts thousands of phishing reports, identifies vulnerabilities before software ships, or cuts a security investigation from hours to minutes is much more interesting.

Once we apply that standard, the market starts to separate quite cleanly.

Alert triage already has strong production evidence. Incident investigation and summarization are useful. Automated response works well when companies tightly define the actions an AI system can take. Vulnerability discovery has recently made a huge jump.

Autonomous penetration testing and autonomous incident response are moving very quickly too, although reliability varies far more between systems.

Cybersecurity AI use	What the evidence looks like now	Our view
Alert triage	Repeated production and controlled-test results	Working well
Incident investigation	Large time savings on routine analysis	Working
Threat detection and correlation	Already deployed at huge scale	Working
Routine automated response	Strong when actions are tightly controlled	Working
Vulnerability discovery	Real production bugs found and fixed	Working surprisingly well
LLM code-security review	Useful but inconsistent	Good second layer
Autonomous penetration testing	Stronger benchmarks, uneven success	Early but real
Autonomous incident response	Real investigations now possible	Emerging fast
Fully autonomous SOC	Little evidence at broad scale	Too early

If you want more recent data on this point, please see our latest cybersecurity market report.

Is AI actually useful for sorting security alerts?

AI alert triage is probably the most proven generative-AI use case in cybersecurity right now.

Microsoft recently published results from randomized testing of its Security Alert Triage Agent. Security professionals using the agent processed user-submitted phishing alerts as much as 78% faster, produced 77% more accurate verdicts and identified 6.5 times more malicious emails.

Those results have also started showing up in production. St. Luke's University Health Network says Microsoft's agent saves its security team more than 200 hours every month by automatically processing large numbers of phishing reports.

A large telecommunications company tested another Microsoft agent on more than 40,000 data-loss-prevention alerts over 90 days. The system filtered that workload down to the roughly 10% of cases considered important enough for human investigation.

CrowdStrike has reported the same basic pattern with Charlotte AI Detection Triage. The company says Charlotte's decisions agree with its Falcon Complete security analysts more than 98% of the time on the alerts it handles.

This type of work suits AI unusually well. The input is repetitive, the system has plenty of previous examples, the decision is relatively narrow, and uncertain cases can still be escalated.

That combination keeps coming up throughout cybersecurity: AI works best when the job has boundaries.

Does cybersecurity AI make analysts better, or just faster?

Cybersecurity AI makes analysts faster today, and the quality improvement is much bigger for less-experienced people.

Microsoft ran two controlled Security Copilot studies that make the difference quite visible.

In its earlier experiment, people with relatively limited security experience completed tasks up to 26% faster with Copilot and produced substantially better results. In the follow-up involving experienced security professionals, analysts were 22% faster but only 7% more accurate.

The speed gain stayed remarkably similar. The accuracy gain did not.

A senior analyst already knows how to interpret a PowerShell script, build a query or structure an investigation. AI mostly removes searching, typing and routine analysis.

A junior analyst has much more knowledge missing from the workflow. The model can supply commands, explain unfamiliar behavior, generate queries and suggest what to check next.

So the biggest workforce effect today may be compression of the experience gap. One experienced analyst can supervise work that previously needed much more direct involvement, while junior analysts reach useful answers faster.

The awkward part comes later. If AI performs more of the simple work that used to train entry-level analysts, companies also need a new way to develop those analysts. SANS's latest workforce research is already picking up that tension, with organizations increasingly saying AI is changing the skills they need rather than simply reducing headcount.

Can AI actually make a SOC respond faster?

AI-heavy security platforms are already cutting response times dramatically, although AI usually works alongside better data integration and automation.

Palo Alto Networks currently has more than 600 active customers using Cortex XSIAM, more than three times the number a year earlier. According to company telemetry and customer interviews, over 60% of deployed XSIAM customers have brought median incident response below ten minutes, compared with days or weeks in their previous setups.

Those customers are collectively feeding more than 15 petabytes of data into XSIAM every day, where Palo Alto says more than 13,000 detection and machine-learning models operate.

Individual deployments help show what sits behind the aggregate.

Konecta, which operates roughly 140,000 endpoints, connected 17 security data sources to XSIAM and reported a 90% reduction in both mean time to detect and mean time to respond. Within four months, 70% of its incident responses had been automated.

Banco Inter reported a 95% drop in detection time and a 98% reduction in resolution time, with 85% of issues closed automatically.

There is an important practical caveat here. These companies also consolidated security tools, cleaned up workflows and automated playbooks. An LLM alone did not produce those gains.

The more interesting result is the system around the model: AI becomes powerful when it already has the company's telemetry and the ability to act.

Deployment	Reported result	What changed underneath
XSIAM customers overall	More than 60% below 10-minute median response	Unified telemetry and automation
Konecta	90% lower detection and response times	17 data sources connected
Banco Inter	95% faster detection, 98% faster resolution	85% of issues closed automatically
Xerox	$10.2M in reported savings	SOC brought in-house and Tier 1 work heavily automated

If you want more recent data on this point, please see our latest cybersecurity market report.

Can cybersecurity AI actually respond to incidents without a human?

Cybersecurity AI can already handle a lot of routine incident response on its own, especially when companies decide in advance exactly what the system is allowed to do.

Konecta says 70% of its incident responses became automated within four months. Banco Inter reports that 85% of issues handled through its XSIAM environment are closed automatically.

That sounds close to a self-running SOC, but the details matter.

Much of this work follows controlled playbooks: enrich an IP address, pull endpoint context, correlate identity events, block a known malicious indicator, isolate a device, update an incident or close a familiar false positive.

The difficult part begins when the system has to interpret an unusual situation and choose an action with a large blast radius.

Automatically quarantining one clearly compromised laptop is relatively easy to approve. Disabling thousands of accounts because an agent misunderstood an identity pattern is a very different decision.

The best current deployments therefore give AI a lot of freedom inside small boxes. Those boxes can already remove enormous amounts of repetitive response work.

Can AI now do a real cyber investigation by itself?

Autonomous cybersecurity agents can now complete surprisingly serious forensic investigations, although the best results depend heavily on the software wrapped around the model.

One of the freshest pieces of evidence comes from SANS's Find Evil competition.

Ninety practicing incident responders ran 1,775 evaluations against 123 finalist systems designed to investigate forensic evidence autonomously. The competition attracted more than 4,400 registrations and 291 working submissions.

The winning agent, Mulder, made 773 logged tool calls while investigating 120 GB of evidence spread across 11 systems. Its workflow planned an investigation, executed forensic tools, analyzed the evidence, challenged its own conclusions and then produced a report.

Another finalist, FindEvil, achieved 98.6% recall across a 552-attack test harness. Protocol SIFT++ refused all 14 destructive actions judges attempted to provoke during testing.

SANS also tested an early forensic agent against a compromised Windows drive. It returned a full analysis in 14 minutes and 27 seconds for work that experienced responders told SANS could normally take days.

The winning systems did more than give a powerful model access to a shell and hope for the best. They used read-only evidence, restricted tools, verification gates, audit logs and separate mechanisms designed to challenge the model's own conclusions.

Autonomous incident investigation is becoming real these days, and the architecture around the AI seems almost as important as the model itself.

Can AI really find security bugs that humans missed?

AI vulnerability discovery is currently producing some of the most convincing evidence anywhere in cybersecurity.

Google's latest Chrome results changed the scale of this argument.

Across Chrome 149 and 150, Google's security teams fixed 1,072 security bugs with help from LLM-based systems. That was more than the 1,036 security fixes shipped across the previous 23 Chrome milestones combined.

Google now uses AI through much of the vulnerability workflow. Models help find bugs, reproduce reports, classify severity, generate candidate patches and create tests. Humans still review the important outputs.

The discovery side has also improved sharply. Google started increasing fuzzing coverage with LLMs several years ago, moved into specialized vulnerability research with Naptime, and then developed Big Sleep with DeepMind and Project Zero. Earlier this year, the Chrome team expanded Gemini-based vulnerability hunting across a much larger portion of the browser's codebase.

One AI-assisted investigation even surfaced a sandbox escape that had remained in Chrome for more than 13 years.

This changes the economics of software security. Humans can inspect code for vulnerabilities, but there are only so many hours available. Agents can repeatedly sweep enormous codebases, revisit old assumptions and test variations that nobody would realistically assign to a security engineer one by one.

The recent Chrome numbers make vulnerability discovery one of the areas where AI's cybersecurity impact is already much bigger than a productivity feature.

Can an LLM replace a normal vulnerability scanner?

LLMs are already useful for finding software-security problems, but traditional scanners still give us consistency that language models struggle to match.

Snyk tested this directly with VulnBench by repeatedly running AI models against the same vulnerable code.

In 300 scans, the strongest high-recall configuration found about 81% of the vulnerabilities contained in Snyk Code's reference set. The best overall configuration reached roughly 75% F1.

The problem was repeatability. Among findings that the models produced outside the scanner's reference set, close to half appeared in only one of five identical runs.

Performance also dropped sharply on the largest, more application-like target. The best model reached only around 40% F1 against the reference findings and repeatedly missed some path-traversal and resource-limit problems.

Still, the LLMs found plausible issues that the conventional scanner had not included in its own results. They were particularly good when a vulnerability depended on understanding what code was trying to do rather than matching a familiar static pattern.

For now, the strongest setup uses both. Deterministic tools scan everything predictably, while an LLM gets another chance to reason across the code and spot strange relationships.

Making the LLM the only security gate would throw away too much consistency.

If you want more recent data on this point, please see our latest cybersecurity market report.

Is autonomous AI penetration testing actually useful now?

Autonomous AI penetration testing has crossed into useful territory, but today's agents are still far from reliably replacing a skilled human pentester.

A recent academic evaluation tested 19 open and proprietary AI models against 300 target servers. The agents received general cybersecurity tools but no target-specific information telling them which service or vulnerability to attack.

Success rates ranged from 10.7% to 69.3%.

The upper end is impressive because the agent had to perform a chain of work: enumerate the target, understand what was running, recognize a possible vulnerability, choose the right tool and successfully exploit the machine.

The spread is just as important as the best score. Two systems presented as “AI pentesters” can currently behave completely differently depending on the underlying model, tools and agent design.

The researchers also found a strong relationship between general model capability and offensive success. That suggests autonomous penetration testing is likely to improve naturally as frontier models improve, even without a major breakthrough specific to cybersecurity.

For repetitive external testing, known vulnerability classes and large attack surfaces, these agents are already useful.

A good human red team still has a major advantage when the job depends on unusual business logic, social context, stealth or improvisation across a messy real company.

Is old-school machine learning still doing most of the cybersecurity work?

Traditional machine learning still carries a huge share of day-to-day cybersecurity today, while generative AI is increasingly handling investigation, reasoning and orchestration around it.

This distinction disappears quickly in vendor marketing.

Security companies have spent years training statistical models to recognize malware, unusual endpoint behavior, suspicious authentication and abnormal network traffic. Those models can continuously score billions of events because the job is narrow and highly optimized.

Generative models add a different capability. They can pull information from several systems, interpret a script, explain why a detection looks suspicious, write a query and decide which tool to call next.

Palo Alto's XSIAM architecture makes the layering easy to see. More than 13,000 detection and machine-learning models continuously process telemetry underneath a newer agent and automation layer. CrowdStrike does something similar by combining Falcon's endpoint data, threat intelligence and existing detection systems with Charlotte AI.

So when a company says “AI stopped an attack,” the useful question is which AI did what.

Quite often, a conventional model detected the suspicious activity, rules and automation gathered the context, and a generative system helped turn the pieces into an investigation.

That mix is currently producing much stronger results than asking one general-purpose model to do everything.

Is AI helping attackers more than defenders?

AI is making cyberattacks easier to scale now, but defenders still have a powerful advantage when they connect AI to private enterprise data and security controls.

IBM's latest Cost of a Data Breach research examined 602 organizations that had experienced breaches. One in four malicious breaches involved AI-enabled techniques, 56% more than the previous year. Those incidents cost an average of roughly $6 million, about $1 million above the overall global breach average.

Attackers are using AI across several stages of the job. Recent threat-intelligence reporting has documented AI-assisted reconnaissance, phishing, malware development, vulnerability research and exploit work.

The autonomous-pentesting benchmark we looked at earlier also shows how quickly that side can improve. A capable model with ordinary security tools can already carry out much of the discovery-to-exploitation chain without being walked through each step.

Defenders have access to something attackers usually lack: the inside of the company.

A security AI system may see endpoint telemetry, employee identities, historical incidents, cloud activity, email, network events and the controls needed to isolate a machine or block an account. Google goes even further by giving defensive AI access to Chrome's source code, test infrastructure and vulnerability history.

That context is a huge advantage when companies actually use it well.

The AI arms race therefore looks much closer than the most dramatic attack stories suggest. Attackers are getting cheaper automation. Defenders can combine the same model improvements with much better data and direct control over the systems being protected.

If you want more recent data on this point, please see our latest cybersecurity market report.

Is cybersecurity AI actually saving companies money?

Cybersecurity AI is already saving real money in some deployments, mostly by shrinking breach duration, analyst workload and the number of security tools companies need to operate.

IBM's latest global breach study found a $1.93 million difference in average breach cost between organizations making extensive use of security AI and automation and organizations using none.

That is large enough to take seriously, although IBM measures AI together with automation, so we should not pretend the entire difference comes from generative models.

Company deployments give us a second type of evidence.

Xerox says its move to an AI-heavy Cortex security operation generated $10.2 million in savings. The company brought its SOC in-house, heavily automated Tier 1 work and later absorbed a security team four times larger during the Lexmark acquisition without rebuilding the old staffing structure.

The Microsoft phishing example is smaller but easier to picture operationally. Saving more than 200 analyst hours every month on one repetitive workflow means roughly 2,400 hours a year can be redirected elsewhere.

Those are the economics that work today. Companies do not need AI to prevent every breach for the investment to make sense. Removing thousands of hours of repetitive work and shortening expensive incidents can already produce a measurable return.

Are AI agents actually replacing cybersecurity jobs?

AI is already replacing parts of entry-level cybersecurity work, while the wider security team is changing more than it is disappearing.

Xerox provides one of the clearest examples. After moving its SOC onto Cortex XSIAM, the company says it eliminated the need for traditional Tier 1 analyst roles.

That is believable because Tier 1 security work contains a lot of exactly what AI handles best: triaging repetitive alerts, collecting context, running known checks, writing summaries and escalating unusual cases.

Microsoft's analyst experiments point in the same direction from another angle. Less-experienced analysts received larger accuracy gains from AI than experienced analysts, while both groups became faster.

So one senior analyst can increasingly supervise a larger amount of work.

The difficult workforce question is what happens to training. Junior analysts historically learned cybersecurity partly by working through large volumes of relatively simple cases. If agents take that layer away, companies still need a path for people to develop the judgment required for harder incidents.

SANS's recent workforce research already reflects this shift. Its latest survey found that 73% of practitioners said AI had changed their team's training requirements, compared with 51% the year before.

The likely outcome is fewer jobs built around moving alerts from one queue to another and more jobs involving investigation, security engineering, AI oversight and unusual incidents.

Why does cybersecurity AI still fail so often if it is this useful?

Cybersecurity AI currently performs extremely well on some narrow jobs and remains frustratingly unreliable once the problem becomes open-ended.

The latest SANS survey captures that contradiction better than almost anything else. As seen above, AI adoption has jumped to 78%, yet 63% of practitioners now report significant shortcomings in AI-based threat detection and response.

Greater usage probably explains part of the increase. More systems in production means more chances to see hallucinations, bad prioritization, incomplete context and strange behavior that a lab test never surfaced.

The newest autonomous-investigation evidence shows how teams are dealing with that problem.

In the SANS Find Evil competition, the strongest entries relied on code-level restrictions, read-only evidence, auditable tool calls and separate checks that tried to disprove the agent's own conclusions. One winning system refused every destructive action judges attempted to trigger.

Google has built similar discipline into Chrome's AI security workflow. Candidate fixes go through testing and human review rather than moving straight into production.

That pattern is one of the most useful lessons in cybersecurity AI today. Reliability improves dramatically when the environment can check the model's work.

Security tasks with hard evidence, tests, logs, restricted permissions or reversible actions are moving fastest. Tasks where “correct” depends on subtle business context remain much harder to hand over.

So what cybersecurity AI is actually working now?

Cybersecurity AI is genuinely working now, especially in alert triage, repetitive investigation, vulnerability discovery, security-data correlation and tightly controlled incident response.

The strongest evidence has moved well beyond chatbot productivity. Microsoft has measured large improvements in phishing triage. Palo Alto customers are reporting incident-response times that have fallen from days to minutes. Google has used AI-assisted security systems to help fix more than a thousand Chrome bugs across two releases. Autonomous forensic agents can now investigate large collections of real evidence, while offensive agents can compromise controlled targets without being told exactly how.

The pattern across all of those cases is remarkably consistent.

AI performs best when the task is specific, the data is good, the tools are already available, and something can check whether the answer makes sense. Companies can give these systems plenty of autonomy inside that kind of environment.

Confidence drops quickly as the job becomes broader and harder to verify.

For that reason, the fully autonomous SOC still looks premature today. Some individual SOC functions are already becoming machine-operated, and Tier 1 work is starting to disappear in the most automated organizations. We have much less evidence that an agent can independently handle the weird, ambiguous and high-stakes incidents that experienced security teams spend most of their attention on.

The stronger conclusion is already clear: cybersecurity AI has found several jobs where machines are plainly better economics than doing everything by hand. The next battle is how much of the difficult tail can follow.

If you want more recent data on this point, please see our latest cybersecurity market report.

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.