Hacking with AI - How Cybercriminals and Ethical Hackers Use Artificial Intelligence in 2025

Executive summary

AI in 2025 is a dual-use force multiplier. Threat actors use AI to increase scale (spear-phishing at mass volume), quality (convincing deepfakes), and speed (automated reconnaissance and triage). Defenders use AI to detect anomalies, prioritize vulnerabilities, and automate incident response.

Key takeaways for CISOs:

• Expect AI to increase volume and sophistication of social engineering (email, chat, voice), and to lower the skill threshold for some financially motivated attackers.
• Defensive investments that reduce human-target risk (phish-resistant MFA, verified wire-transfer processes, behavioral analytics) produce the strongest short-term ROI.
• Adopt AI governance and vendor assessments: logging, provenance, and model-use policies are now essential.
• Prepare for model-assisted malware and synthetic-media threats; collaborate with peers and threat intelligence communities to share indicators quickly.

Why 2025 is different

Three structural changes make 2025 distinct:

1. Democratization of capability: Off-the-shelf LLMs, accessible audio/video synthesis, and agent frameworks reduce the technical barrier for sophisticated scams.
2. Commoditization of “attack building blocks”: Underground markets and illicit services now sell AI-enabled phishing kits, voice-clone plugins, and reconnaissance-as-a-service.
3. Vendor intervention and visibility: Large AI providers are actively disrupting misuse, publishing threat reports, and building mitigations — which both reduces impact and reveals novel abuse patterns.

Threat taxonomy: AI-enabled attacks (high-level)

Below is a non-exhaustive, high-level classification of AI-related threats. Each entry intentionally avoids operational details and focuses on risk and detection vectors.

1. AI-enhanced social engineering

Personalized messages that combine public data, behavioral signals and natural language generation to craft highly convincing lures across email, SMS, Teams/Slack, and social media.

2. Deepfake vishing and multimedia impersonation

Voice cloning and synthetic video used to impersonate executives or customers for fraud or extortion. Public reporting shows a measurable surge in voice-based scams and impersonation attempts.

3. AI-assisted reconnaissance & prioritization

Attackers use models to summarize large corpora (code, commit histories, public cloud metadata) and prioritize likely attack paths — reducing time-to-exploit.

4. Model-assisted malware / polymorphism (proofs-of-concept)

Research and limited incident data show attackers exploring AI-assisted code synthesis for obfuscation and polymorphism; academic work and vendor reporting highlight evolving proofs-of-concept. This remains an area of active research and monitoring.

5. Disinformation & coordinated influence

Automated content generation used in targeted influence operations; AI amplifies content creation and message variation. Vendor reporting documents interdicted campaigns.

6. Low-skill criminal automation

“Attack-as-a-service” bundles and marketplaces reduce the need for programming skill — enabling more actors to run scams at scale.

Representative incidents & research

OpenAI public threat disruption reporting (2024–2025)

OpenAI published reports describing detection and disruption of abusive behavior including influence campaigns and social engineering leveraging AI. These reports illustrate both the misuse and the vendor mitigation posture.

KELA 2025 AI Threat Report

KELA’s 2025 report documents rising mentions of malicious AI tools and provides detailed case examples of how underground actors repurpose AI for phishing, reconnaissance, and financial fraud.

Academic proofs-of-concept

Academic research has demonstrated how an LLM can orchestrate attack stages in lab settings — highlighting feasibility concerns. Some projects simulated AI-orchestrated ransomware as a warning sign for the future.

Deepfake vishing trend reporting

Security vendors and media outlets have tracked increases in voice-clone enabled scams and executive impersonations, particularly impacting finance and HR teams.

Risk matrix & prioritization (for CISOs)

Use this compact matrix to focus investments. Columns: Likelihood, Impact, Priority, Example mitigations.

Detection & response playbook (CISO-oriented)

This playbook is prescriptive at the control and process level (not procedural malware steps).

1) Short-term (0–30 days) — Reduce human-target risk

• Enforce phish-resistant MFA (FIDO/WebAuthn) for all privileged roles and payment approvers.
• Implement or update wire-transfer and invoice approval procedures: multi-person verification and out-of-band confirmation.
• Run focused phishing simulations using AI-like content to test terrain (with consent and legal clearance).

2) Medium-term (30–90 days) — Strengthen detection and telemetry

• Deploy and tune behavioral EDR/XDR analytics for lateral movement and anomalous process behavior.
• Integrate email and messaging telemetry with SIEM/XDR to correlate cross-channel signals.
• Ingest threat-intel feeds that include AI-related IOCs and join trusted sharing groups.

3) Longer-term (90+ days) — Governance, provenance and resilience

• Adopt AI governance policies and vendor risk assessments.
• Invest in media provenance tools such as watermarking and metadata validation.
• Run purple-team exercises that simulate AI-assisted attack scenarios.

Runbook examples

• Trigger: CFO receives urgent voice call requesting fund transfer. Action: Treat as high-risk, require in-person or token-based confirmation, escalate to fraud desk.
• Trigger: Multiple employees click on a convincing email. Action: Disable affected accounts, reset MFA, investigate for lateral movement.

Technical controls & reference architectures

Recommended defensive stack

• Phish-resistant authentication such as FIDO2/WebAuthn for privileged and payment roles.
• Multi-layer email defense: MTA controls, DMARC/SPF/DKIM, and AI/behavioral analysis.
• EDR/XDR with behavioral detection across endpoints, network, and cloud.
• SOAR automation for triage with human approval for critical actions.
• Media provenance verification for sensitive communications and executive media.

Architecture note

Do not over-automate high-impact security decisions. Keep a human in the loop for financial actions and major remediation steps.

Data & telemetry priorities

1. Authentication logs and MFA events
2. Email and collaboration platform telemetry
3. Endpoint process and network flow data
4. Cloud configuration and IAM events

AI governance, procurement & vendor risk

AI governance should be part of enterprise security. Recommendations include:

• Define model-use policies (e.g., no sensitive data in public prompts).
• Require vendors to provide security controls, red-team summaries, and disclosure commitments.
• Log employee use of AI tools for business decisions and maintain provenance for official media.
• Ensure contracts allow for forensic cooperation if abuse is detected.

People, training, and purple team exercises

Human factors are the strongest defense against AI-powered social engineering.

• Provide realistic employee training including phishing, vishing, and deepfake awareness.
• Run purple team exercises to test how well detection and response work against AI-generated attacks.
• Brief executives and boards using business-focused language: financial loss, reputation, regulatory exposure.

CISO 90-day practical roadmap

1. Days 0–30: Deploy phish-resistant MFA, review wire-transfer approvals, and run a tabletop exercise on deepfake fraud.
2. Days 31–60: Integrate telemetry across email, endpoints, and cloud; subscribe to AI threat intelligence; run a purple team drill.
3. Days 61–90: Formalize AI governance, require vendor red-team reports, pilot provenance tools, and update incident playbooks.

Frequently Asked Questions

Is AI making cybercrime inevitable?

No. AI increases scale and speed for some threats, but layered defenses and governance reduce risk significantly.

Are voice-deepfake scams widespread?

Yes. They are rising and often target finance teams with impersonation calls and urgent requests.

Can attackers use LLMs to write malware autonomously?

Research proofs-of-concept exist, but real-world attacks still rely on human direction alongside tools.

What is the fastest defensive win?

Deploy phish-resistant MFA and strengthen transaction authorization processes.

Should companies ban LLM use internally?

No. Governance policies and monitoring are more effective than outright bans.

How do you detect synthetic audio?

Use voice-provenance analysis, monitor call patterns, and require secondary verification.

Do signature-based tools still matter?

Yes, but they should be complemented by behavioral analytics to counter AI-based evasion.

Can threat intelligence help with AI threats?

Yes. AI-related threat intel feeds provide trends, IOCs, and campaign details that aid defenses.

What training should organizations run?

Phishing simulations, vishing awareness campaigns, and payment-approval drills.

Are AI vendors cooperating on security?

Yes. Many vendors publish misuse reports and are building safeguards. Enterprises should require transparency.

Will regulation solve the issue?

Regulation helps, but organizations must act proactively. Laws take time to catch up with fast-moving threats.

What is adversarial machine learning?

It is the study of how manipulated inputs can fool AI models. Security teams should test resilience against it.

What should you do if you suspect a deepfake fraud?

Escalate immediately, preserve evidence, notify fraud teams, and involve law enforcement.

How important is media provenance?

Very important for verifying authenticity of sensitive or executive-level communications.

Should organizations use AI for detection?

Yes, but always with human oversight for high-impact decisions.

How do you prioritize AI risks?

Start with financial fraud and credential theft, then address disinformation and reputational risks.

Are nation-states using AI in cyber operations?

Yes. Reports confirm exploration of AI for influence operations and automation.

Can small businesses defend effectively?

Yes. Implement MFA, regular backups, phishing training, and verified payment approvals.

What are indicators of AI-enabled campaigns?

High-volume personalized messages, rapid variation of content, and synthetic media across channels.

Where can I learn more?

Review public threat reports, industry intelligence briefs, and academic research on adversarial ML.

References

• OpenAI: Reports on disrupting malicious AI use
• KELA: 2025 AI Threat Report
• ThreatLabz / Trellix: AI and cybersecurity research publications
• Media outlets reporting on deepfake vishing trends
• Academic publications on adversarial machine learning