Claude Sdk For Python EUVD-2026-17677

The CVSS score of 4.8 with vector AV:L/AC:L/PR:L reflects a local attack requiring low privileges and no user interaction, resulting in low confidentiality and integrity impacts. The EPSS score is not provided in the input data. The vulnerability is NOT listed in CISA KEV and no public exploit code is identified, indicating this is a lower-urgency issue from an exploitation perspective despite the clear security defect. However, real-world risk varies significantly by deployment context: in shared hosting environments or multi-tenant systems, local users could exfiltrate sensitive agent state or inject malicious instructions into AI model interactions, creating indirect but serious impact. In containerized deployments (highlighted as a specific attack vector), the risk is elevated if containers run with permissive umask or if multiple containers share a volume with world-writable permissions. The low CVSS score reflects the requirement for local access; however, the specific nature of the exposure-sensitive AI agent state and model behavior manipulation-warrants prompt patching despite the low numerical risk rating.

On a shared Linux host or Docker container with a permissive umask, a local attacker running as a different user (or from another container with shared volume access) could enumerate the home directory or application data directory of a process using the affected SDK version, discover the memory files, read them to extract sensitive prompt histories and model instructions, and potentially modify them to inject adversarial instructions into subsequent AI model invocations. For example, in a multi-tenant container platform where volumes are shared, an attacker in one tenant's container could read or corrupt agent memory belonging to another tenant, compromising the confidentiality and integrity of AI-driven applications.