Cyber Risks · The Hacker News

AI Flaws in Amazon Bedrock, Lang Smith, and SGLang Enable Data Exfiltration and RCE

Summary: BeyondTrust reported that Amazon Bedrock AgentCore Code Interpreter's sandbox mode allows outbound DNS queries that can be used to build command-and-control channels, obtain interactive reverse shells, and exfiltrate data from AWS resources the service's IAM role can reach; the issue has no CVE and carries a CVSS score of 7.5. Amazon, notified in September 2025, classified the behavior as intended functionality and recommends VPC mode and a DNS firewall, while separate disclosures covered a URL parameter injection flaw in LangSmith (CVE-2026-25750, CVSS 8.5, fixed in version 0.12.71) and unpatched deserialization vulnerabilities in SGLang (CVE-2026-3059 and CVE-2026-3060) that can lead to remote code execution.

Why it matters: A sandbox sold as isolated still let DNS traffic out, and an overprivileged IAM role is what turns that into access to S3 buckets holding sensitive data. SGLang's flaws remain unpatched, so any deployment exposing its multimodal generation or ZMQ broker port to the network is reachable by an unauthenticated attacker.

Read the original at The Hacker News ↗

Topics
Cloud Security
Mentions
CVE-2026-25750
Related