Cognitive　radio　(CR)　is　regarded　as　the　key　technology　of　the　6th-Generation　(6G)　wireless　network.　Because　6G　CR　networks　are　anticipated　to　offer　worldwide　coverage,　increase　cost　efficiency,　enhance　spectrum　utilization,　and　improve　device　intelligence　and　network　safety.　This　article　studies　the　secrecy　communication　in　an　energy-harvesting　(EH)-enabled　Cognitive　Internet　of　Things　(EH-CIoT)　network　with　a　cooperative　jammer.　The　secondary　transmitters　(STs)　and　the　jammer　first　harvest　the　energy　from　the　received　radio　frequency　(RF)　signals　in　the　EH　phase.　Then,　in　the　subsequent　wireless　information　transfer　(WIT)　phase,　the　STs　transmit　secrecy　information　to　their　intended　receivers　in　the　presence　of　eavesdroppers　while　the　jammer　sends　the　jamming　signal　to　confuse　the　eavesdroppers.　To　evaluate　the　system　secrecy　performance,　we　derive　the　instantaneous　secrecy　rate　and　the　closed-form　expression　of　secrecy　outage　probability　(SOP).　Furthermore,　we　propose　a　deep　reinforcement　learning　(DRL)-based　framework　for　the　joint　EH　time　and　transmission　power　allocation　problems.　Specifically,　a　pair　of　ST　and　jammer　over　each　time　block　is　modeled　as　an　agent　which　is　dynamically　interacting　with　the　environment　by　the　state,　action,　and　reward　mechanisms.　To　better　find　the　optimal　solutions　to　the　proposed　problems,　the　long　short-term　memory　(LSTM)　network　and　the　generative　adversarial　networks　(GANs)　are　combined　with　the　classical　DRL　algorithm.　The　simulation　results　show　that　our　proposed　method　is　highly　effective　in　maximizing　the　secrecy　rate　while　minimizing　the　SOP　compared　with　other　existing　schemes.