jenkins.security.Security218Test#jnlpSlave fails not infrequently on Windows as in e.g. this run. I suspect this is due to the split of instance-identity. Based on the logs of that run and other similar runs, Jenkins#getAgentProtocols must be returning the empty set, and I suspect the reason why is that the initialization of JnlpSlaveAgentProtocol4 (which depends on instance-identity having been initialized) is racing with the initialization of instance-identity. Our Windows test agents are very slow, so the problem can probably be reproduced by adding a sleep statement somewhere.