Configuration-Embedded Credential Extraction
Security & IdentityDescription
API keys, passwords, database credentials, and authentication tokens embedded in agent configurations are extracted by adversaries. Agents often require credentials to access their tools, creating high-value credential concentration.
Attacker gains access to agent's system prompt and extracts AWS access keys hardcoded for S3 access.
Prompt injection and inference-time manipulation have been extensively demonstrated against production LLM applications in red-team exercises and public disclosures, with researchers successfully hijacking agents via malicious content in retrieved documents. Confirmed large-scale production incidents causing measurable financial or safety harm remain limited.
Primary mitigations
- Secrets management systems (not config files)
- credential rotation policies
- agent config encryption
- runtime secrets injection
- credential access auditing.
Detection signals
Credential exposure events in agent configurations; secrets management compliance rate.