Key takeaways
- AI is now embedded across the cybersecurity stack — threat detection, endpoint security, email filtering, identity analytics, and incident response.
- Defensive AI works by learning what normal looks like and flagging deviations, using anomaly detection, classification, and graph analysis.
- Offensive AI — attackers using generative AI to craft phishing, malware, and social engineering — is rising quickly and raising the baseline threat level.
- The MITRE ATT&CK framework structures how defenders map detections; AI increasingly automates coverage across its tactics and techniques.
- Human analysts remain essential. The winning pattern is AI for speed and scale, humans for judgment and context.
Why AI matters in security
The scale of modern threat landscape is beyond human processing. A mid-sized company’s SIEM (security information and event management) platform ingests billions of events per day. A single ransomware operator can launch thousands of simultaneous campaigns. Static rule-based detection misses novel attacks. Statistical and machine-learning approaches are the only path to triaging signal from noise at this scale.
The underlying technique is our anomaly detection primer — find the events that don’t fit learned patterns. Around that core, cybersecurity has built layered approaches specialized for specific threat classes. For broader ML context, see our machine learning coverage.
Defensive AI use cases
Endpoint detection and response (EDR)
Modern EDR products (CrowdStrike Falcon, SentinelOne, Microsoft Defender for Endpoint, Palo Alto Cortex XDR) use ML on every endpoint to detect malicious process trees, fileless attacks, and living-off-the-land techniques. The industry moved from signature-based antivirus to behavioural ML around 2015-2018; signature AV is now considered a supporting detection, not the primary one.
Network detection and response (NDR)
NDR platforms analyze network traffic for command-and-control beacons, lateral movement, and data exfiltration. Supervised classifiers flag known-bad patterns. Unsupervised anomaly detection catches novel traffic. Graph analysis traces attack paths across many hops.
Email and phishing defense
Email gateways use NLP to detect phishing, business email compromise, and payload-bearing attachments. Microsoft Defender for Office 365, Proofpoint, Abnormal Security, and Mimecast all embed ML classifiers. Deep-learning content analysis combined with sender-reputation graphs has pushed phishing detection to high accuracy, though sophisticated targeted phishing still gets through.
Identity and access analytics
User and entity behaviour analytics (UEBA) profiles normal behaviour per account and flags anomalies — a user logging in from two continents in five minutes, accessing systems they have never used, pulling data at unusual volume. Okta, Microsoft Entra, Splunk UBA, and Exabeam are representative.
SOC automation (SOAR)
Security Orchestration, Automation, and Response platforms use AI to triage and enrich alerts before a human sees them. LLMs are increasingly used to summarize alerts, draft initial investigation notes, and recommend containment actions. Palo Alto Cortex XSOAR, Splunk SOAR, and IBM QRadar SOAR lead this category.
The offensive AI threat
Attackers use AI too. Three categories are visible in the wild:
AI-crafted phishing
Generative AI produces convincing phishing content in any language, with personalized details scraped from LinkedIn and public sources. Microsoft’s 2024 Digital Defense Report flagged a substantial uptick in AI-generated phishing, much of it harder to detect than traditional phishing because of better grammar and contextual awareness.
Deepfake social engineering
Voice cloning from a few seconds of audio. Video deepfakes of executives. Multiple high-profile incidents in 2024 involved attackers impersonating CFOs on video calls to authorize fraudulent wire transfers. The Hong Kong Arup case, where a finance employee was tricked into transferring $25 million after a deepfake video call, illustrates the stakes.
Automated exploit generation and malware
AI can rewrite malware to evade signatures, generate polymorphic variants at scale, and identify vulnerabilities in code. Concerns about AI-generated zero-day exploits remain partly theoretical at the frontier — current models are not yet reliably producing novel working exploits — but the trajectory is clear. Defensive research expects this capability to mature within years.
Supply-chain and model security
As AI becomes part of every product, the AI itself becomes an attack surface. Poisoned training data, compromised pre-trained models from Hugging Face or GitHub, adversarial examples that cause misclassification, and prompt-injection attacks on LLMs all need defensive consideration. The OWASP Top 10 for LLM Applications (2023, updated 2024) catalogs the main risk categories. For broader safety context, see our ai safety coverage.
What works in practice
ML-assisted triage beats pure ML detection
Alert volume is the existential problem. Fully autonomous detection produces too many false positives. The durable pattern: use ML to score and cluster alerts, put the highest-confidence cases in a priority queue, use lower-confidence cases for human-driven investigation. This reduces analyst fatigue without giving up coverage.
Defense in depth
No single AI model catches everything. Effective programs layer detections across endpoints, network, identity, email, and cloud — each running its own ML with distinct assumptions. When an attacker evades one, another catches them. This is classic defense-in-depth applied to AI-augmented controls.
Human-in-the-loop for high-stakes responses
Auto-remediation (kill process, quarantine host, disable account) is risky. An overzealous response takes critical systems offline. Most mature programs gate automated responses behind thresholds — below a score, auto-contain; between scores, alert; above a score, require human approval.
Organizational impact
Cybersecurity workforce demand continues to exceed supply — the US alone has several hundred thousand unfilled security jobs. AI does not close this gap; it shifts what roles are needed. Analysts move up the stack from first-level triage (increasingly automated) to threat hunting, incident response, and red-team work. Programs that deploy AI without upgrading analyst skills see less benefit than ones that integrate the two.
Regulatory and frameworks
Cybersecurity AI deployment is shaped by compliance frameworks — SOC 2, ISO 27001, NIST Cybersecurity Framework, HIPAA for healthcare, PCI-DSS for payments. The CISA Secure by Design initiative and NIST’s AI Risk Management Framework are increasingly referenced. EU AI Act classifications will affect how AI in critical infrastructure and security products is developed and documented.
Frequently asked questions
Can AI replace security analysts?
No, and the companies that have tried have mostly backtracked. AI is excellent at scoring, clustering, and enriching alerts, but the final investigation — connecting dots, judging whether a detection matches a real incident, communicating with stakeholders, making containment decisions — remains human work. The successful pattern is AI as force multiplier for a smaller, more skilled analyst team rather than AI-only SOCs.
Should I use an AI-powered antivirus over a traditional one?
Modern endpoint security products all use ML now — the line between “AI-powered” and “traditional” antivirus has blurred. The meaningful choice is between next-generation EDR/XDR platforms and older signature-focused antivirus. For any business beyond the smallest, an EDR or XDR product is much more effective. For home users, the built-in Microsoft Defender and Apple’s protections are generally adequate; third-party consumer antivirus has declined in necessity.
Are AI-powered phishing attacks worse than before?
The floor has risen. Bad grammar and awkward phrasing — historically signals of phishing — are no longer reliable tells, because LLMs write clean text in any language. Personalization quality has improved. On the other hand, defensive AI has improved in parallel, and the volume of AI-generated phishing has itself become a detectable signal in aggregate. The arms race continues, and the net effect is that cybersecurity investment and user training are more important, not less.
