Hackers Bypass AI Safeguards Using Claude, Deepseek — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- CrowdStrike researchers analyzed four threat groups using agentic AI between May and August, finding all had compromised targets by circumventing model safeguards with older versions of Anthropic’s Claude and Deepseek models.
- Chinese-speaking hackers used a runtime patch called ClawGod to strip 'CYBER_RISK_INSTRUCTION' from Claude Code and targeted U.S. and Brazilian organizations in biotech, healthcare, and tech sectors.
- Spanish- and English-speaking attackers built a penetration testing harness using Claude Opus 4.5, Haiku 4.5, and Deepseek V4 Pro to autonomously attack Argentine healthcare entities, stealing credentials and medical records from at least one victim.
- Russian- and Indonesian-speaking actors used basic AI tools but still compromised government domains in Indonesia and Bangladesh, showing even low-sophistication groups can succeed with minimal infrastructure.
- Adam Meyers of CrowdStrike noted attackers exploit differences in model guardrails by routing tasks across AI systems, often framing malicious activity as authorized security testing to bypass restrictions.
- Multiple threat groups left their AI infrastructure exposed online, allowing researchers to observe how they used supervisor and executor agent architectures to run long-term, autonomous hacking operations.
Why it matters: Organizations face rising AI-powered threats from both sophisticated and low-skilled hackers who can now automate attacks at scale using widely available models, undermining the assumed security of AI guardrails and increasing the risk of large-scale data breaches in critical sectors like healthcare and government—where defenses often fail basic hygiene.
Ask SkimNews
