Base 64 Encoded Content Detected In AI Logs
Updated: August 19, 2026
Description
Base64-encoded data has been found in AI logs.
Base64 itself is not harmful, its presence may hide sensitive or unexpected information-such as secrets, tokens, or entire documents-that bypasses regular log filters or monitoring tools.
Base64 content in logs may indicate attempts to obfuscate data, misuse of the system, or improper data handling. It can also lead to compliance issues if sensitive content is being stored without proper controls.
Example Attack
An internal user sends a prompt to an AI model containing a Base64-encoded file for processing. The full encoded string is logged by the system. Later, security teams discover the string contains a decoded PDF with customer financial data, unintentionally exposing confidential information in the logs. This creates both a security risk and a compliance violation.
Remediation
Review the AI logs to understand what content is being encoded and why. If Base64 data is not expected, update input validation and logging rules to detect and block encoded payloads. Inspect the source of this data and ensure sensitive content like API keys, credentials, or files is not being transmitted or logged in this format.
Security Frameworks
Inputs that could steer model behavior must be normalized, screened by a prompt injection detection ruleset or classifier, and subordinated to a system-level instruction hierarchy.
Automated classifiers must scan every response and block harmful content, system prompt disclosure, outbound request triggers, and hidden or encoded content.
Known jailbreak patterns, prompt injection attempts, adversarial inputs, probing behavior, and granular token usage must be detected, attributed, and alerted on.