GoKawiilCrowdStrike finds four hacker groups routinely bypassing AI model safeguards
CrowdStrike researchers analyzed AI infrastructure from four previously unreported threat groups operating between May and August, writing in Chinese, Spanish/English, Indonesian and Russian. Two built sophisticated AI 'harnesses' while two less technically advanced groups still compromised multiple targets, mostly relying on older versions of Anthropic's Claude and open-weight Deepseek models left exposed on public servers.
GoKawiil's interpretation of the reporting above, not reported fact.
The findings suggest that attackers don't need cutting-edge AI to automate intrusions, since cost-conscious use of older models and switching between systems can let them exploit uneven guardrails across providers. CrowdStrike's Adam Meyers frames model routing as a deliberate tactic, implying that AI safety protections may need to be consistent across vendors to have real effect against low-resource adversaries.
- Four distinct hacker groups used agentic AI setups left openly accessible, letting CrowdStrike study their methods directly.
- Attackers favored older, cheaper AI models over frontier ones, mirroring cost-saving patterns seen in legitimate AI use.
- Inconsistent guardrails across different AI models let attackers switch tools to evade safety restrictions.
YubiKey 5 Series Security Key — As AI-powered hacking campaigns grow more autonomous and harder to detect, hardware-based multi-factor authentication is one of the few defenses that doesn't depend on guessing whether an attacker is a bot or human. A YubiKey adds a physical layer of protection to your accounts that phishing kits and automated credential-stuffing tools can't easily bypass.》
See YubiKey 5 Series Security Key on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.Source: wired.com — Lily Hay Newman, 2026-10-21
Published there as: “AI Guardrails Aren’t Stopping Criminal Hackers”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.