🟠 High | Source: AWS Security Bulletins
A flaw in the Strands Agents Tools package (versions prior to 0.8.2) allows an attacker to steal credentials configured for HTTP requests by manipulating the proxy settings via a crafted prompt. The hostname allowlist intended to restrict where credentials are sent remains satisfied, but the credential is leaked to an attacker-controlled proxy in cleartext. This is particularly dangerous in agentic AI applications that process untrusted external content, making indirect prompt injection a realistic attack path.
Security Architect’s Take: Upgrade strands-agents-tools to version 0.8.2 or later immediately. Review any deployed agents that use the http_request tool with HTTP_REQUEST_TOKEN_CONFIG and audit logs for unexpected outbound proxy connections that may indicate prior exploitation.
Original advisory: CVE-2026-18394 - Incorrect authorization in Strands Agents Tools http_request tool