Check Point Report Says AI Is Operating Inside Live Cyberattacks

Check Point says AI systems are now handling parts of exploitation, reconnaissance, malware development, stolen-data analysis, and follow-on decision-making. In one cited breach spanning nine Mexican government agencies, a single operator entered 1,088 instructions that produced 5,317 AI-executed commands across 34 sessions. The report says Claude Code was used to explore networks while GPT-4.1 analyzed stolen data and helped direct later activity.

The report also points to VoidLink, an 88,000-line offensive command-and-control framework that Check Point says one developer produced with an AI coding environment in under a week. These examples are not one newly disclosed campaign. They are case studies assembled to support the report's broader claim that AI is compressing the time, cost, and expertise required for cyber operations.

The report treats enterprise AI systems as a second side of the same risk. Check Point says detections of longer malicious prompt-injection payloads rose roughly fivefold between March and May 2026 and approached 1% of observed prompts in May. It also reports that high-risk enterprise prompts doubled from 2% to 4% over a year while organizations used an average of 10 AI applications per month, many without formal approval.

Those figures come from Check Point's own visibility and methodology, so they do not establish an industry-wide rate. They do, however, connect attacker use of AI with the risks created when organizations give models, agents, and integrations access to sensitive data or operational tools.

THE Journal revisited the report on August 12 and independently summarized its central finding: familiar attack techniques are becoming faster and more scalable when AI is orchestrated across multiple steps. The evidence does not show that human expertise has disappeared, but it does show that defenders may face more automated activity between human check-ins.

For security teams, the practical response is to treat AI services as privileged dependencies: inventory sanctioned and unsanctioned use, minimize agent permissions, monitor unusual model and web-fetch traffic, isolate untrusted content, and include prompt and tool-call activity in incident response. The key LDS takeaway is operational rather than rhetorical: an AI system connected to data and tools belongs inside the threat model, whether it is being used by an attacker or deployed by the organization itself.