It’s tempting to think AI-driven attacks need AI-driven defenses. New research from CrowdStrike, covered by WIRED, suggests something more mundane: attackers are using AI to hit the same old weaknesses, just faster and at greater volume.
What CrowdStrike found
- Researchers studied four previously unreported groups active between May and August, working in Chinese, Spanish and English, Indonesian and Russian.
- All four left their AI setups exposed on web servers, which is how researchers could see what they did. They mostly used versions of Anthropic’s Claude and open-weight DeepSeek models, usually older versions rather than the newest.
- All four got around model safeguards, often by persuading the model the work was an authorized penetration test, including by planting files meant to “prove” approval.
- The Spanish- and English-speaking group compromised at least 15 healthcare organizations in Argentina using a “supervisor” agent directing an “executor” agent.
- Even the least sophisticated group compromised government domains.
CrowdStrike’s Adam Meyers told WIRED these attacks weren’t using new exploits; they exploited security hygiene problems that “were preventable but weren’t prevented.” He also noted a human was usually doing initial probing before handing off to AI.
Field notes: the fix list
- Patch internet-facing systems first. Automated attackers scan everything. Your public servers, VPNs and web apps should be on the fastest patch cycle you have.
- Enforce MFA everywhere, especially on remote access, email and admin accounts. Stolen credentials featured in these campaigns.
- Remove default and stale accounts. Old test logins and unused admin accounts are exactly what bulk scanning finds.
- Watch for speed. AI-run attacks move quickly and steadily. Alert on bursts of login attempts, rapid directory scans and unusual volumes of requests from one source.
- Reduce what’s exposed. Every admin panel or test site on the public internet is a target. Put them behind a VPN or access proxy.
- Use AI on defense too. The same tools can help you triage alerts and review configurations faster.
The takeaway
Model guardrails help, but they aren’t a security control you can rely on. The good news in CrowdStrike’s findings is that the defenses that work are ones you already know: patching, MFA, fewer exposed systems and alerting on unusual activity.
Based on reporting by WIRED and CrowdStrike.
