v1.4.8
AIPentest/CyberStrikeAIv1.4.8Apr 3, 2026by Ed1s0nZ
AI Summary
The multi-agent attack chain generation logic has been improved to correctly identify tool usage by analyzing process details and filling incomplete ReAct inputs.
Key Highlights
- Attack chain generation now relies on process_details in addition to mcp_execution_ids.
- Fills incomplete ReAct inputs using the last assistant turn's details.
- Filtering logic separates main agent tools from sub-agent details.
New Features
- Enhanced multi-agent attack chain generation
- Process detail analysis for tool detection
- ReAct input completion logic
Full Release Notes
多代理模式下，攻击链原先只认助手消息里的 mcp_execution_ids，容易因缺少 execution id 被误判为「无工具」而无法生成。现已改为同时依据 process_details 中的 tool_call/tool_result 判定；并在 ReAct 保存不完整时，用最后一轮助手的过程详情补全输入，且按主编排器视角过滤——保留主 agent 的工具调用与子任务调度（task）、子 agent 仅保留最终回复，去掉 thinking/planning 及子 agent 工具细节，使攻击链生成与「主 agent 整轮迭代」一致。 --- In multi-agent mode, attack-chain generation previously required MCP execution IDs on assistant messages, so runs without execution IDs were treated as having no tools and produced empty chains. The fix also treats persisted tool_call/tool_result events as evidence of tool use, and when saved ReAct input is incomplete, appends filtered process details for the last assistant turn: orchestrator tools and task dispatches are kept, sub-agents contribute only their final replies, while thinking/planning and sub-agent tool traces are dropped—aligning the prompt with one full orchestrator iteration.