The Role of Artificial Intelligence (AI) in Modern Cybersecurity

A phishing email written by a large language model gets a 54% click-through rate, more than four times the rate of a hand-written lure. On the defensive side, the same class of technology now flags account takeovers and insider threats in minutes instead of days. Artificial intelligence (AI) has become the fastest-moving variable in cybersecurity, and it sits on both sides of the fight at once.

Published: December 2023. Updated: August 2026. Reviewed by Encryption Consulting’s Security Engineering and Cryptographic Advisory team.

This is not a narrow cryptography question. AI in cybersecurity spans detection engineering, security operations, fraud and social engineering, and a growing intersection with encryption and public key infrastructure (PKI), since AI models, agents, and the data pipelines behind them all need an identity, and identity in enterprise security still runs on certificates and keys. The rest of this article covers both directions: what AI does for defenders, what it does for attackers, and where it touches the cryptographic controls Encryption Consulting builds for a living.

How Is AI Changing Cybersecurity Defense?

AI changes defense by finding patterns across data volumes no analyst team can review manually, then acting on those patterns faster than a human response cycle allows. Four capabilities account for most of the real gains.

Anomaly and behavioral detection (UEBA)

User and entity behavior analytics (UEBA) models build a baseline of normal login times, data access volumes, and network paths for every account, then flag deviations. This is how AI catches an insider threat or a compromised credential that never trips a signature-based rule, because the login itself looks legitimate; only the behavior around it is wrong.

Automated triage and response

Security operations centers (SOCs) use AI to correlate alerts across SIEM, EDR, and cloud logs, cut duplicate and low-fidelity noise, and recommend or trigger a first response, such as isolating a host or disabling a session. This is the single biggest lever on mean time to respond (MTTR), which is the metric that actually determines how much damage a breach does.

Phishing and social engineering detection

Email security platforms now use language models to score message intent, sender behavior, and linguistic patterns rather than relying only on known-bad links or attachment hashes. This matters more every year because the phishing they are screening for is itself increasingly AI-written.

Code and vulnerability scanning

AI-assisted static and dynamic analysis tools scan source code and dependencies for known weakness patterns faster and with fewer false negatives than purely rule-based scanners, helping engineering teams catch flaws before release rather than after an incident.

How Is AI Changing the Attacker’s Toolkit?

AI lowers the cost and raises the quality of attacks that used to require real skill or manual effort. Three trends matter most for security teams building a 2026 threat model.

Deepfakes and synthetic identity fraud

Voice and video deepfakes are now routinely used in business email compromise and vendor payment fraud, where an attacker impersonates a CFO or vendor contact on a live call. Pindrop measured voice deepfake fraud attempts against contact centers rising from about one per month to roughly seven per day during 2024, and Deloitte’s Center for Financial Services projects generative AI-enabled fraud losses in the United States could reach $40 billion by 2027, up from $12.3 billion in 2023 (Deloitte). Human accuracy at spotting video deepfakes unaided sits close to chance in independent testing, which is why detection has to be technical, not just trained vigilance.

AI-generated phishing and business email compromise

Generative models now write grammatically clean, context-aware phishing at scale, personalized from scraped LinkedIn data and breached mailing lists. The FBI’s Internet Crime Complaint Center (IC3) recorded roughly $20.9 billion in reported cybercrime losses for 2025, with business email compromise and phishing-adjacent fraud among the largest categories (FBI IC3 2025 Annual Report). The defensive answer is the same AI-driven email filtering described above, kept current against a lure that keeps improving.

Adversarial machine learning

Adversarial ML attacks target the AI system itself: data poisoning corrupts a training set so the model learns the wrong pattern, evasion attacks craft inputs that slip past a classifier, and prompt injection manipulates an LLM’s instructions through its own input. MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) is the standard reference here, cataloging these techniques the way MITRE ATT&CK catalogs conventional attack techniques; as of early 2026 it documents 16 tactics and 84 techniques specific to AI and ML systems (MITRE ATLAS). Any organization deploying AI for security should assume its own models are a target, not just a tool.

For a deeper look at one of these attack classes specifically, see Encryption Consulting’s guide to prompt injection attacks and the Education Center’s breakdown of generative AI threats and mitigations.

Where Does AI Meet Encryption and PKI?

AI and cryptography intersect in three concrete ways: AI helping secure cryptographic infrastructure, cryptography securing AI systems, and AI models themselves needing the identity and key-management controls every other production system already has.

AI-assisted certificate and key anomaly detection

Machine learning models applied to certificate transparency logs, certificate authority (CA) issuance records, and key usage telemetry can flag misissuance, unauthorized certificate authority behavior, weak or duplicate keys, and unusual key access patterns well before a manual audit would catch them. This is the same anomaly-detection logic described earlier, applied to certificates and keys instead of user sessions.

AI-driven crypto-agility and cryptographic inventory

You cannot secure, migrate, or automate what you have not inventoried. Tools like Encryption Consulting’s CBOM Secure use automated discovery and classification, increasingly AI-assisted, to build a cryptographic bill of materials (CBOM) across an enterprise: which algorithms, key lengths, certificates, and libraries are in use, and where. This inventory is the prerequisite for crypto-agility, meaning the ability to swap a weak or deprecated algorithm without rebuilding the system around it, and it is exactly the ground truth an AI-driven security tool needs to reason correctly about cryptographic risk. An AI model trained on an incomplete inventory will miss the assets it never saw.

Key management for AI models and agents

Autonomous AI agents that call APIs, move data, or take action on a security team’s behalf need a verifiable identity, the same way a service account or a machine does. That identity is typically a short-lived certificate issued and rotated through PKI, not a static API key or shared secret, because a static credential handed to an autonomous agent is a durable target. See Encryption Consulting’s guide to securing AI-to-AI communication with private PKI for the mechanics of applying mutual TLS (mTLS) and certificate-based identity to agent-to-agent, agent-to-model, and orchestrator-to-worker traffic.

Choosing AI-Driven Security Tools: What to Evaluate

Not every product marketed as “AI-powered” uses the same technique, and the technique determines what the tool is actually good at. Evaluate against these criteria before buying.

• Model type and fit

  Supervised classifiers work well for known attack patterns with labeled training data (malware families, known phishing kits). Unsupervised anomaly detection fits novel or insider threats where there is no labeled example. Large language models fit natural-language tasks such as alert summarization and phishing content analysis, not deterministic classification. A vendor should be able to say plainly which category their model falls into.

• False positive and false negative rates, stated with numbers

  Ask for a benchmark against your own log samples, not a marketing claim. A high false-positive rate causes alert fatigue and gets the tool ignored; a high false-negative rate is worse and harder to detect.

• Explainability

  Can the tool show why it flagged something, in terms an analyst can act on? A black-box score with no reasoning slows investigation and makes the tool hard to defend in an audit or a post-incident review.

• Integration and data format compatibility

  Confirm it ingests your existing SIEM, EDR, and cloud log formats without a custom pipeline, and that its output maps to your SOAR playbooks. A tool that requires reformatting every data source adds operational cost that erodes the time savings.

• Data residency and training data handling

  Know whether your logs and telemetry train the vendor’s shared model, stay in a tenant-isolated instance, or leave your environment at all. This is a contractual and compliance question, not just a technical one.

• Human-in-the-loop controls

  Confirm the tool supports a detection-only or approval-gated mode before it supports autonomous action, and that the gate is actually enforced, not just available.

A Balanced Threat Model: AI Security Use Cases, Value, and Risk

A useful threat model treats AI as dual-use: every deployment carries defensive value and a corresponding risk that needs a control. The table below maps common use cases against both.

AI security use case	Defensive value	Primary risk
UEBA / anomaly detection	High: catches unknown-pattern threats and insider risk	Medium: false positives, model drift as normal behavior changes
SOC alert triage and automated response	High: cuts mean time to respond	Medium to high: autonomous action needs strict guardrails
AI-driven phishing and email filtering	High: catches AI-generated lures signature tools miss	Low to medium: adversarial evasion, novel lure formats
AI-assisted code and vulnerability scanning	Medium to high: finds flaws faster, pre-release	Medium: false negatives on novel logic flaws
AI-assisted certificate and key anomaly detection	Medium to high: surfaces misissuance and key misuse early	Medium: accuracy depends entirely on inventory completeness
Generative AI copilots for security analysts	Medium: speeds investigation and documentation	Medium to high: prompt injection, data leakage into prompts
Autonomous AI agents with system access	Medium: extends coverage beyond human hours	High: expands attack surface, needs strong agent identity and key management

How to Adopt AI Security Tools: A Step-by-Step Process

Most failed AI security deployments fail at rollout, not at the model. Follow this sequence rather than turning autonomous features on by default.

1. Inventory your data sources and, where relevant, your cryptographic assets first, since an AI tool is only as good as the ground truth it reasons over.
2. Define the threat model and success metrics for the specific use case before evaluating vendors, rather than buying a general “AI security platform.”
3. Pilot the tool in detection-only or shadow mode against real production traffic before granting any response authority.
4. Tune for false-positive rate against your own environment and integrate its output with your existing SIEM and SOAR playbooks.
5. Add explicit human-in-the-loop approval gates before allowing any autonomous or irreversible response action.
6. Extend your existing identity and key-management controls to cover the AI system itself, including any agent identities it creates or uses.
7. Monitor for model drift and re-evaluate detection accuracy on a fixed schedule, at minimum quarterly, since attacker behavior and your own environment both change.

Performance and Interoperability Trade-Offs

AI security tools carry real operational costs that a proof-of-concept demo rarely surfaces.

• Inference latency

  Real-time inline detection (blocking a request before it completes) needs low-latency models, typically classical ML rather than large language models. Batch or near-real-time analysis can tolerate a heavier model and higher accuracy.

• Compute and cost at scale

  Running large language models across every log line or every email is expensive; most production deployments apply the LLM selectively, after a cheaper classical model has narrowed the field.

• Data normalization overhead

  Feeding a model consistent, well-labeled data from a dozen disparate tools (firewalls, EDR, cloud audit logs, certificate authorities) is usually the largest hidden cost of a deployment, not the model itself.

• Vendor lock-in on training data

  A model trained heavily on one vendor’s proprietary telemetry format can be difficult to port if you switch SIEM or EDR providers later.

Deployment Examples

A few concrete patterns show how this plays out in production rather than in the abstract.

• Bank fraud detection

  A retail bank runs UEBA against transaction and login telemetry to flag account takeover attempts in real time, cutting the window between compromise and detection from days to minutes.

• SOC alert triage across hybrid cloud

  An enterprise SOC uses an AI copilot to correlate alerts from on-premises AD CS logs, cloud audit trails, and EDR telemetry into a single prioritized queue, reducing analyst alert volume by an order of magnitude.

• Cryptographic inventory and crypto-agility

  CBOM Secure applies automated discovery and classification across certificates, keys, and libraries to build a live cryptographic inventory, the foundation both for catching cryptographic anomalies and for planning a post-quantum migration.

• Enterprise email security

  An organization deploys AI-based email filtering specifically tuned against AI-generated phishing patterns, alongside standard sender authentication controls (SPF, DKIM, DMARC).

• Autonomous agent identity

  A DevOps team issues short-lived, PKI-backed certificates to autonomous AI agents that call internal APIs, so each agent action is cryptographically attributable and revocable rather than authenticated by a long-lived shared secret.

Limitations

AI is not a substitute for a security program, and it has real, well-documented failure modes.

• False positives and false negatives

  Every model has an error rate. A false negative on a targeted attack is the costliest failure mode and the hardest one to catch, since by definition the model did not flag it.

• Model drift

  A model trained on last year’s baseline behavior degrades as the environment, user population, and attacker tactics change, and needs retraining or re-tuning on a schedule, not a one-time deployment.

• Explainability gaps

  Deep learning models in particular can produce a risk score without a reviewable reason, which slows investigation and can be a compliance problem in regulated industries.

• Adversarial risk to the model itself

  Data poisoning, evasion, and prompt injection, cataloged in MITRE ATLAS, target the AI system directly. A security tool that is itself unhardened against these techniques can become the weak point it was deployed to close.

• Automation complacency

  Teams that trust automated triage too completely stop reviewing the edge cases a model was never trained to catch, which is exactly where a sophisticated, targeted attack lives.

• Data privacy and leakage

  Feeding sensitive logs, code, or customer data into a generative AI copilot without clear data handling guarantees can create a new exposure the organization did not have before.

What Would Encryption Consulting Recommend?

Start from inventory, not from the AI tool. Most organizations that struggle to get value from AI-driven security products are missing the cryptographic and asset ground truth the model needs to reason correctly, which is why CBOM Secure exists as a discovery and classification layer before any crypto-agility or anomaly-detection tool is trusted to make decisions. Pair that with an Encryption Advisory engagement to govern the crypto-agility roadmap the inventory reveals, and see how managing cryptographic posture as a continuous discipline, not a point-in-time audit, keeps that inventory accurate as your environment changes.

Second, treat every AI agent as a machine identity from day one. An autonomous agent calling internal systems needs the same short-lived, certificate-backed identity and revocation path as any service account, not a static API key that becomes a standing liability the moment it leaks. Encryption Consulting’s PKI-as-a-Service and private PKI offerings extend that identity model to agent-to-agent and agent-to-model traffic, detailed in the AI-to-AI communication guide referenced above.

Third, keep a human approval gate on anything above low-risk automated action, and re-evaluate that gate as the tool’s track record accumulates rather than removing it on day one because the vendor demo looked confident.

Conclusion

AI has moved from an experimental add-on to a core part of both cybersecurity defense and the modern attacker’s toolkit, and 2026 is not the year to treat either side of that as optional. Defensively, AI is genuinely strong at anomaly detection, automated triage, and catching phishing that has itself gotten harder to spot by eye. Offensively, deepfakes and AI-written social engineering have already produced billions of dollars in measured fraud losses, and adversarial ML gives attackers a documented, growing playbook against the models defenders now rely on.

Start with the inventory question before the tool question. If you cannot say with confidence what certificates, keys, and algorithms are running in your environment today, that gap will limit every AI-driven security investment you make on top of it.

References

• NIST, AI Risk Management Framework (AI RMF 1.0), released January 26, 2023, and NIST-AI-600-1, Generative Artificial Intelligence Profile, released July 26, 2024.
• MITRE, MITRE ATLAS, adversarial threat landscape for AI systems, 16 tactics and 84 techniques as of early 2026.
• FBI Internet Crime Complaint Center, 2025 IC3 Annual Report.
• Deloitte Center for Financial Services, Generative AI is expected to magnify the risk of deepfakes and other fraud in banking.
• Pindrop, 2025 Voice Intelligence and Security Report, cited via industry analysis on contact center deepfake fraud growth.