JustLog3 Cloud's AI assistant reads logs and repositories, creating a prompt-injection vulnerability where crafted log entries can manipulate model outputs. Red-team testing found two exploitable attack vectors: links that exfiltrate data from logs, and web search results that poison the assistant's memory with malicious instructions. Fixes include sanitizing URLs against workspace data and blocking memory saves from external sources.
AWS AgentCore's security is compromised by a prompt that requests credentials, enabling potential infrastructure attacks. The headline appears among broader tech news covering AI system vulnerabilities, semiconductor developments, and cybersecurity incidents.
A comprehensive guide comparing 10 prompt-injection detection and defense tools for production AI agents across five control layers: red-team testing, production detection, guardrails, gateway controls, and action authorization. Tools range from cloud-native managed services like Azure Prompt Shields and Bedrock Guardrails to self-hosted options like Llama Prompt Guard 2 and open-source frameworks like garak and PyRIT, with evaluation criteria covering deployment location, indirect injection detection, agent workflow support, and action-control mechanisms.
The article argues that prompt injection remains an unsolved problem in AI systems and dismisses claims of solutions as either dishonest or indicative of misunderstanding the fundamental challenge.
Zenity researchers discovered a critical vulnerability in AWS Bedrock AgentCore that allowed attackers to hijack all AI agents in an AWS account through a single prompt to one public agent. The flaw stemmed from improper isolation and overly broad default permissions, enabling access to credentials, source code, and private data. AWS has partially patched the issue but recommends companies implement stricter access controls.