In the Mexican operation, Claude Code and OpenAI’s GPT-4.1 were used to break in, move laterally, and analyze data, all controlled by a single operator.

The report finds most enterprise data exposure comes from ordinary, approved use, as employees share extra context to get useful AI results rather than from overt external attacks.

Identity cannot be trusted as the sole security control in an era of synthetic media, since AI can generate convincing voices, faces, and documents, with human reviewers identifying only about 41% of AI-generated faces.

Prompt injection remains a growing threat, with longer malicious payloads rising fivefold from March to May 2026 and approaching 1% of observed prompts in May.

AI is now producing ready-to-deploy malware and attack frameworks, such as an 88,000-line C2 framework built with an AI coding environment in under a week.

Check Point’s 2026 AI Security Report warns that AI has advanced from aiding attacks to executing live intrusions, with cases tied to China-linked espionage and a multi-agency Mexican government breach.

A single operator ran two commercial AI tools in parallel, conducting live intrusions across nine Mexican government agencies and generating 5,317 AI-executed commands over 34 sessions.

Experts argue that regulation alone won’t be enough; effective defense requires technical controls, user awareness, and continuous monitoring of AI tool use.

High-risk enterprise prompts have doubled, rising from about 1 in 50 to 1 in 25 interactions, as organizations average ten AI applications per month and a large majority report at least one high-risk interaction monthly.

Vulnerability remediation windows are shrinking to potentially as little as 12 hours on critical internet-facing systems as regulators push for ultra-rapid response to AI-enabled vulnerability windows.

Risk exposure varies by industry, with Business Services showing the highest high-risk prompt rate, signaling uneven security controls and usage patterns across sectors.

AI expands the attack surface through data processing influences, surrounding software vulnerabilities, and supply-chain risks, with model hubs and agent skills showing widespread weaknesses.