bartek@aws: ~/news
$ whoami

Bartek Chojnacki

$ AWS Architect · DevOps · Cloud
Tuesday, August 4, 2026

Amazon Bedrock AgentCore Security Flaw Lets Users Bypass AI Safety Guards

Amazon Bedrock AgentCore's InvokeHarness API has a nasty input validation bug (CVE-2026-18830) that lets authenticated users trigger configured tools without the AI model checking them first. If your harness has tools set up, attackers could invoke them directly, skipping all those safety controls you've got in place. The good news? Impact is limited to whatever tools you've actually configured—empty harnesses are safe. You'll want to update immediately if you're running versions before July 31, 2026.

source: [aws/security-bulletin]

Kiro IDE and CLI on Windows: Critical Executable Resolution Vulnerability

Kiro IDE (versions 1.0.0–1.0.212) and Kiro CLI (pre-2.10.0) on Windows have a nasty path traversal issue—CVE-2026-18656 and CVE-2026-18657. A malicious project directory can plant an executable that runs before your system PATH, letting attackers execute arbitrary code when you open the folder. **Action required**: Update immediately if you're on Windows.

source: [aws/security-bulletin]

Amazon Bedrock Now Lets GPT Models Search the Web Without Leaving AWS

Amazon just dropped Web Search on Bedrock, and it's a game-changer for keeping your data locked down. Instead of juggling third-party search APIs and compliance reviews, you can now ground GPT responses with current web knowledge using a single API parameter—all server-side within AWS. The tech leverages Amazon's massive web index and knowledge graph to deliver semantic snippets optimized for your model's context window, complete with citations. No vendor onboarding, no data egress, no headaches.

source: [aws/whats-new]

also that day:

Monday, August 3, 2026

Strands Agents Tools: Prompt Injection Bypasses Shell Command Approval

Strands Agents Tools versions below 0.8.0 have a critical vulnerability (CVE-2026-18733) where prompt injection can bypass the shell tool's consent gate. An attacker can craft prompts—like those hidden in untrusted content—to set the non_interactive parameter to true, letting arbitrary OS commands execute without operator approval. Update to version 0.8.0 or later immediately if you're running affected versions.

source: [aws/security-bulletin]

AWS CLI EMR Commands Have SSH Verification Flaw – Update Required

AWS CLI versions 1.45.27 and earlier (v1) or 2.35.2 and earlier (v2) have a nasty security gap: the EMR SSH helper commands (aws emr ssh, aws emr socks, aws emr put, aws emr get) skip host key verification. This opens the door for man-in-the-middle attacks to intercept your SSH sessions and file transfers. You need to update immediately to patch CVE-2026-18654.

source: [aws/security-bulletin]

AWS Amazon MQ MCP Server: Critical Credential Leak via Prompt Injection

AWS Amazon MQ MCP Server versions up to 2.0.23 have a nasty vulnerability (CVE-2026-18655) that lets attackers snag your broker credentials and OAuth tokens through prompt injection. The flaw affects RabbitMQ connection tools—basically, a crafted endpoint can trick the system into leaking sensitive auth data. You need to upgrade to version 2.0.24 or later immediately. This one's serious and requires immediate action.

source: [aws/security-bulletin]

also that day:

Saturday, August 1, 2026

Master Cost and Performance Trade-offs with Amazon EMR Advanced Scaling

Amazon EMR on EC2 just got smarter with Advanced Managed Scaling—pick your strategy based on what matters most: squeeze costs with conservative scaling, balance both worlds, or go aggressive for SLA-critical jobs. The post walks you through real scenarios so you can dial in the perfect scaling config for your data pipeline.

source: [aws/big-data-blog]