CVE-2026-49844: Log4j API: NaN/Infinity breaks JSON log serialization

Improper encoding of non-finite floating-point values during MapMessage JSON serialization in Apache Log4j API produces output that is not valid JSON. This issue affects Apache Log4j API versions 2.13.1 through 2.25.4 and version 2.26.0. The fix for CVE-2026-34481 did not cover all code paths: when a MapMessage contains a non-finite IEEE 754 value (NaN, Infinity, or -Infinity), MapMessage.asJson() emits the corresponding bare token. RFC 8259 does not permit these tokens, so a conformant parser rejects the resulting document. The defect is reachable only when both of the following conditions hold: * The application uses the message resolver https://logging.apache.org/log4j/2.x/manual/json-template-layout.html#event-template-resolver-message of JsonTemplateLayout or any other layout that relies on MapMessage.asJson() or MapMessage.getFormattedMessage(new String[]{"JSON"}). * The application logs a MapMessage that contains an attacker-controlled floating-point value. An attacker who can supply a non-finite value can cause the affected layout to emit malformed JSON, which may corrupt the enclosing log record or disrupt downstream log ingestion and parsing. Users are advised to upgrade to Apache Log4j API 2.25.5 or 2.26.1, both of which emit RFC 8259-compliant JSON for non-finite values.

An adversary identifies a JVM-based service (e.g., a model-serving API or MLOps component) that logs a MapMessage containing a numeric field they can influence — for example, a model confidence score, latency metric, or telemetry value submitted via an API parameter. By supplying a value that resolves to NaN or Infinity (e.g., via a division-by-zero condition or an out-of-range input), the attacker causes JsonTemplateLayout's message resolver to emit a bare NaN/Infinity token in the log's JSON payload. This corrupts the log record, potentially breaking the downstream log ingestion pipeline or SIEM parser, disrupting security monitoring, or — in more sophisticated cases — being used defensively by an attacker to blind automated detection tooling that depends on parsing that log stream, effectively masking other malicious activity occurring around the same event.