As the pace of adopting and advancing artificial intelligence (AI) has outstripped the capabilities of existing security safeguards (guardrails), more than half of corporations are facing risks from new types of cyberattacks, an analysis found.

Akamai Technologies, a cybersecurity and cloud company, said in its report published on the 10th, "2026 corporate AI use risk report," that unapproved "shadow AI," hyper-enabled advanced AI users, and stealthy browser extensions are exposing corporations' core asset to new types of cybersecurity risks (threats).

Akamai researchers found three new AI-based attack techniques this year: Vibe hacking, CursorJacking, and CometJacking.

Vibe hacking is an attack that secretly tampers with directive files in the development environment that AI coding tools reference, inducing the AI to create vulnerable code or perform tasks the attacker wants.

CursorJacking is an attack that uses malicious browser extensions to covertly collect sensitive information—such as API keys, source code, and chat logs—that popular AI coding tools like Cursor access.

CometJacking is an attack that hides malicious commands on public web pages to control AI agents and exfiltrate users' files, emails, and login information, targeting agent-based browsers such as Comet AI from Perplexity.

Or Eshed, Akamai vice president of enterprise security product and engineering institutional sector, said, "AI is no longer just a productivity tool; it has established itself as a collaborative partner with direct access to corporations' core asset (crown jewels)."

The problem is that existing data loss prevention (DLP) tools were built for an era centered on file transfers and email, Eshed noted.

He emphasized, "Today, sensitive corporate data is being systematically fragmented across millions of fluid prompts, unmanaged personal accounts, and autonomous AI agents," adding, "Security leaders must move away from approaches that simply try to block AI and shift toward continuously controlling and managing how AI operates at the interaction layer level."

※ This article has been translated by AI. Share your feedback here.