Anthropic has published details confirming that Russian-linked threat actors used Claude to automate malware evasion - using the model to iteratively refine malicious code and reduce its detectability. That alone is a significant data point on how capable AI tools are being folded into attack workflows.
But the more notable finding is the second part of the disclosure: the same campaign involved a direct intrusion attempt against Anthropic's own infrastructure, targeting a pre-release version of Claude. AI vendors are no longer just a conduit for misuse - they are themselves a high-value target.
The implication for security teams is worth sitting with. If the organisations building these models can't be assumed safe from sophisticated intrusion, then third-party AI dependencies carry supply chain risk in the same way any other vendor relationship does. That belongs in your risk register, not just your acceptable use policy.
Source: SecurityWeek