Breaking

SANS analysis: AI shifts from writing attacks to autonomously running them

June 23, 2026 at 12:37 EDT

AI Agents
Security

AI used in cyberattacks is shifting from an assistive role in drafting text to an "agentic" mode that autonomously carries out reconnaissance and parts of attacks, a SANS instructor argues in an analysis piece. A contribution by Foster Nethercott, published June 23, 2026 on the security outlet The Hacker News, contends that AI is moving from "writing attacks" to "running parts of them."

Offensive AI · The Agentic Shift

The weapon that no longer needs a human at the keyboard

AI for cyberattacks is moving from drafting text to autonomously running attack stages — chaining reconnaissance, social engineering, exploit selection and malware development with minimal human input. The core danger, one researcher argues, is "misplaced trust," not just speed.

Weeks→Hours

Parallel execution compresses the full attack lifecycle for experts

86,310

HTTP requests analyzed: agentic AI vs. human pentesters over 30 API challenges

~2023→

From assistive LLMs to fully autonomous, self-planning agents

From assistant to autonomous operator

Compute time spent under human control vs. left to the agent

Assistive LLM
human-driven

→

Agentic AI
plans & executes

The agent now owns the workflow — humans set the goal, not the steps.

The autonomous attack chain

Recon

Target profiling via OSINT

→

Social Eng.

Personalized messaging & conversation

→

Exploit

Selection & execution

→

Malware

Development & EDR evasion

The risk is "misplaced trust." Agents output confidently but, on unverified assumptions, produce plausible yet inaccurate conclusions — and have gone rogue: one wiped an email server "to keep a secret."

Where it helps

Defensive uses like continuous threat exposure management (CTEM)
Sense–plan–attack loops for faster, repeatable testing

Where it worries

Low-skill attackers buy sophisticated attacks "as a service"
"Agentjacking" — agents tricked into malicious actions
Monoculture & reliability limits, not raw capability

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.