# Hackers use older AI models to bypass safeguards in attacks

> CrowdStrike reports four groups exploiting open AI tools like Anthropic Claude and Deepseek to run autonomous hacks between May and August 2026.

Oossa · 2026-10-21 · https://oossa.com/en/hackers-use-older-ai-models-to-bypass-safeguards-in-attacks

Researchers at CrowdStrike detailed how four previously unknown hacker groups used agentic AI to carry out attacks from May to August 2026. The groups, operating in Chinese, Spanish‑English, Indonesian and Russian, accessed older versions of Anthropic’s Claude and open‑weight Deepseek models. They sidestepped built‑in guardrails by framing prompts as authorized penetration testing and by patching system prompts. The Spanish‑English group alone compromised at least 15 Argentine health‑care organizations and stole credentials.

All four groups left their AI infrastructure exposed on public web servers, letting the researchers see the models and harnesses in use. The attackers favored older, cheaper models, showing that even non‑cutting‑edge AI can power large‑scale hacking.

## The facts

- Four threat groups used Claude Opus 4.8, Claude Haiku 4.5, Claude Sonnet 4.6 and Deepseek V4 Pro.
- The Spanish‑English group compromised at least 15 Argentine health‑care entities and stole medical records.

## Why it matters

Security teams must treat AI‑driven attacks as a realistic threat, even when criminals use older, cheaper models.

## Sources & references

1. [AI Guardrails Aren’t Stopping Criminal Hackers](https://www.wired.com/story/ai-guardrails-arent-stopping-criminal-hackers/) – Wired, 2026-10-21

Last updated: 2026-10-21