AI-Powered Cyberattacks

AI-powered cyberattacks use machine learning or language models to make parts of an attack faster, cheaper or more convincing. Agentic cyberattacks go further: an AI agent plans and executes several attack steps with limited human direction. For defenders, the practical question is not whether AI was used, but what the attack looks like in the evidence and how to investigate it.

Two Levels of AI Involvement

	AI-assisted	Agentic (more autonomous)
What AI does	Writes phishing text, translates, summarizes stolen data, helps write code	Chains reconnaissance, exploitation and lateral movement with minimal human input
Human attacker role	Directs every step	Sets goals and supervises
Effect for defenders	Better-quality lures, more volume	Faster tempo, many parallel attempts, rapid adaptation

Public reporting from AI providers and security vendors through 2025 and 2026 has documented both, with AI-assisted use far more common. Claims of fully autonomous attacks should be read carefully. See Agentic AI for how agents work in general.

What Defenders Can Observe

AI does not leave a stamp on network packets. Defenders see the same categories of evidence as always, but the patterns can change:

• Identity: High volumes of well-written, personalized phishing; rapid password spraying adapted to responses; MFA fatigue at scale.
• Endpoint: Scripts that vary between hosts; quick retries with modified commands after a failure; heavy use of built-in tools.
• Network: Many short, systematic reconnaissance probes; unusually fast movement between discovery and exploitation.
• Tempo: Steps that would take a human hours compressed into minutes, often around the clock.

Investigating a Suspicious Sequence (Controlled Synthetic Scenario)

Clearly labelled synthetic scenario run in a test environment.

Time	Observation
02:10:04	Internal host DEV-17 begins querying directory services for all users and groups
02:10:41	SMB connections attempted to 212 hosts in 37 seconds
02:11:30	Failed exploit attempt against FILE-02; second attempt with a different payload 9 seconds later
02:11:52	Successful authentication to FILE-02 with a service account
02:12:40	Archive created of \\FILE-02\finance

Reconstruction: Discovery → mass probing → exploit retry with variation → credential use → staging. About 2.5 minutes in total.

Competing explanations

• Vulnerability scanner: rejected; no scanner is assigned to DEV-17 and scanners do not create archives.
• Human operator with scripts: possible; scripted tooling can be this fast.
• Agentic tooling: possible; the adaptive retry and tempo are consistent with it.

What the evidence establishes: A malicious intrusion and data staging from DEV-17. What it does not establish: Whether a human, a script or an AI agent chose each step.

What Evidence Can and Cannot Show About AI Use

• Can show: speed, parallelism, adaptive variation and the full sequence of actions.
• Cannot reliably show: that a language model or agent was involved, unless you recover the attacker's tooling, prompts or infrastructure.
• Implication: Investigate and respond based on what happened. Attribution to AI is rarely needed to contain an incident and is often impossible to prove.

Defensive Priorities

1. Shrink response time. Faster attacks require faster investigation; this is where automated, explainable investigation helps most. See Automated Threat Investigation.
2. Detect sequences, not single events. See Behavioral Threat Detection.
3. Harden identity. Phishing-resistant MFA removes a large share of AI-boosted phishing value.
4. Keep humans on approvals, with case files that let them decide quickly.

Key Takeaways

• Most AI use by attackers today is assistance; agentic attacks raise tempo and adaptation.
• Defenders observe the same evidence types, with faster and more varied patterns.
• Investigate what happened; proving AI involvement is usually neither possible nor necessary.

Related reading: AI Safety and Alignment, Explainable AI in Cybersecurity and Microsoft 365 Account Takeover.