Researchers at Palo Alto Networks‘ Unit 42 just showed how a few hidden words buried inside an ordinary support ticket could trick an AWS AI agent into handing over live credentials, and AWS’s response afterward is arguably the more interesting part of the story.
The team targeted AgentCore Harness, the AWS infrastructure layer that gives AI agents access to tools, memory and external services. Using a simulated customer support agent connected to an external server, researchers planted hidden instructions inside a support ticket, a classic indirect prompt injection. The manipulated agent triggered the Harness’s built-in shell tool, which runs as root by default and, crucially, could read memory belonging to its own runtime process.
That access turned out to matter a great deal. When the Harness authenticated to an external service using a credential stored through AWS’s identity system, that credential briefly existed in plaintext inside process memory, exactly where the compromised shell could read it.
Researchers pulled out a usable authentication token, sent it to a simulated external server they controlled, and then successfully reused that token from an entirely separate system, no AWS credentials required. From there, they reached a customer lookup tool containing test personal data.
AWS reviewed Unit 42’s disclosure and closed it as informational, pointing to its shared-responsibility model. Under that framework, AWS handles things like microVM isolation and infrastructure security, while customers remain responsible for input validation, tool restrictions and monitoring what happens inside their own agent sessions. In other words, AWS built the house, but it’s telling customers the locks on individual rooms are their job to install.
That framing tracks with a broader pattern researchers keep running into. NIST’s Center for AI Standards and Innovation ran similar tests on agents with command-line access and found they could frequently be manipulated into executing malicious instructions or leaking data, independent of any specific vendor’s platform.
Unit 42’s practical advice lands squarely on the customer side of that shared-responsibility line: restrict which built-in tools each agent session can actually use, limit the privileges assigned to service accounts, and watch outbound network traffic closely. None of that is exotic security advice. It’s just easy to skip when a default configuration works fine until the day someone slips instructions into a support ticket.

