There are two bugs:
    1. It looks like the at command does not validate that it has succeeded to write all data before it makes the file executable (i.e on a disk full an empty file might get created).
    2. After 60 minutes (run_time + CHECK_INTERVAL <= now), atd (approx line 750) starts to unlink the lockfile and retries the execution of the job, leaving us with this unfortunate loop.
    
    Reproducer:
    ATD_PID=$(ps -C atd -o pid=)
    QUEUE=a
    JOBNO=$(printf '%05x' 123)
    BAD_TIME=$(expr $(date +%s) / 60 - 61)
    CTM=$(printf '%08x' $BAD_TIME )
    FILENAME=/var/spool/at/${QUEUE}${JOBNO}${CTM}
    touch $FILENAME
    chmod 0700 $FILENAME
    kill -HUP ${ATD_PID}
    sleep 0.5
    ls -l $FILENAME
    rm -f $FILENAME
    
    Thanks to: Anders Blomdell