The cybersecurity landscape has fundamentally shifted as attackers now deploy artificial intelligence not just as an advisory tool, but as an autonomous operator.
This structural transformation has reduced the time required to discover and exploit zero-day vulnerabilities from months to mere minutes.
Until late 2025, cybercriminals primarily used AI as an assistant to draft phishing messages or generate simple code snippets.
Today, AI models independently scan networks, attempt exploits, and dynamically adjust tactics using standards such as the Model Context Protocol.
AI-powered offensive frameworks have demonstrated that complex vulnerabilities can be successfully exploited in under 10 minutes. This automation drastically lowers the skill barrier, allowing a single threat actor to match the output of an entire security team.
To address this expanding attack surface, tracking frameworks like MITRE ATLAS have rapidly updated their tactics and techniques to categorize these new automated capabilities.
The first documented autonomous AI espionage operation, attributed to the state-backed group GTG-1002, compromised several global entities in late 2025.
Attackers bypassed safety guardrails by exploiting a task-decomposition jailbreak on Claude Code, enabling the AI to manage the operation with minimal human intervention.
During this campaign, the model successfully conducted autonomous discovery, privilege escalation, and zero-day exploitation in a live environment.
State-sponsored actors like APT28 have also utilized large language models in operations such as the LAMEHUG campaign.
This malware queried cloud-based AI models in real time to generate malicious commands, highlighting how threat actors test dynamic, on-the-fly capabilities.
Dynamic Malware and Defense
The integration of AI into malicious payloads has given rise to a new category of dynamic, self-modifying malware. Programs like MalTerminal use AI endpoints to generate ransomware or reverse shell payloads directly in memory at runtime.
Experimental droppers like PROMPTFLUX continuously rewrite their source code by querying external APIs to apply obfuscation techniques that evade traditional antivirus software.
Other tools include PROMPTLOCK, which uses local AI models for ransomware scripting, and QUIETVAULT, which manipulates legitimate AI tools on a target machine to steal secrets.
Threat actors weaponize these capabilities so rapidly that enterprise networking infrastructure is facing an all-time high of zero-day exploits.
Beyond traditional malware cyberthint, financially motivated actors have utilized AI for pure data exfiltration and extortion in large-scale “vibe hacking” operations against various institutions.
However, this reliance on AI introduces the hallucination paradox into the threat ecosystem. Cases like the React2Shell incident revealed that some AI-generated exploits are entirely hallucinated and non-functional.
While AI tools accelerate attack speeds and decrease lateral movement times to mere minutes, their tendency to hallucinate creates a unique vulnerability.
Follow us on Google News , LinkedIn and X to Get More Instant Updates. Set Cyberpress as a Preferred Source in Google.
