Check Accounts After AI-Driven Hackers Bypass Safety Guardrails in Recent Campaigns

Organizations and their employees face increased exposure to automated cyberattacks that leverage artificial intelligence to bypass standard safety controls. Recent analysis by security firm CrowdStrike identifies four previously unreported threat groups that operated between May and August. These actors targeted common security hygiene issues within victim environments, exploiting speed and volume rather than discovering new software flaws.

Root cause

The attacks succeeded because the threat groups manipulated agentic AI tools to automate malicious tasks. According to CrowdStrike, the attackers did not rely on novel exploits or zero-day vulnerabilities. Instead, they targeted basic security hygiene issues that were preventable but remained unaddressed in many environments. The researchers noted that the attacks leveraged the speed and scale that AI provides, allowing criminals to conduct operations at a volume that overwhelms traditional defenses.

Attack path

The threat groups operated in Chinese, Spanish, English, Indonesian, and Russian languages. The Chinese, Spanish, and English-speaking groups exhibited sophisticated AI harnesses that guided models to carry out tasks. The Indonesian and Russian-speaking groups showed less technical prowess but still compromised multiple targets. All four groups left their AI infrastructure openly accessible on web servers, which allowed researchers to analyze their activity. They used this infrastructure to run campaigns autonomously over relatively long periods.

Affected versions

The attackers primarily used versions of Anthropic’s Claude and open-weight Deepseek models. According to CrowdStrike senior vice president Adam Meyers, the groups favored older model versions over frontier models to manage costs. The actors used both open- and closed-weight models. They selected specific models for different operational aspects, routing requests to take advantage of varying guardrails. This strategy allowed them to bypass protections that differed across models, even from the same company.

Mitigation

• Review access logs for autonomous or automated activity that does not match normal business hours.
• Ensure that AI infrastructure and associated web servers are not left openly accessible.
• Address common security hygiene issues such as weak credentials and unpatched systems.
• Monitor for rapid, high-volume operations that may indicate automated tooling.

For customers and users

Teams usually recommend checking account activity for unusual access patterns or locations. Employees should verify that their credentials have not been reused across other services. Watch for follow-up scams that reference this incident to steal more data. Ensure that multi-factor authentication is enabled on all critical accounts. Report any suspicious emails or messages that claim to be from IT support asking for login details. Regular password changes remain a basic defense against automated credential stuffing attacks.

What to do and how to stay safe: CrowdStrike

• Audit your organization's web servers to ensure AI infrastructure is not publicly accessible.
• Enforce strong password policies and multi-factor authentication to mitigate credential-based attacks.
• Monitor network traffic for high-volume automated requests that may indicate agentic AI usage.
• Review security hygiene practices to address common vulnerabilities that attackers target for speed.

Step-by-step guide: Passkeys Explained: How Passwordless Login Actually Works

General security guidance from the DarkWeb Headlines newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Are hackers using new vulnerabilities to bypass AI guardrails?

No, the attacks targeted common security hygiene issues rather than novel exploits or vulnerabilities.

Which AI models were used in these attacks?

The attackers primarily used versions of Anthropic’s Claude and open-weight Deepseek models.

Did the attackers use frontier AI models?

No, they mostly used older model versions to manage costs and avoid overkill.