Breakdown of HackAgent's 19 attack techniques: AI Agent jailbreak testing and prompt injection

The open-source red-team tool HackAgent v0.17.0 includes 19 jailbreak and prompt injection attack techniques, and the same YAML configuration can be rerun with a different model. The author installed 0.17.0 locally and actually ran the validate and dry_run commands. The attacks are divided into 9 static, 8 adaptive, and 2 multi-turn, of which four — rag, stac, tool_output_ipi, and toolhijacker — have their entry point in the data the Agent reads rather than in the conversation. On the evaluation side, multi-judge voting is supported, with majority as the default and a threshold of 7.0 (out of 10); when a judge call fails or the answer cannot be read, it votes abstain rather than safe, and if all abstain it is recorded as error and not included in the denominator. An optional audit checks the accuracy of each judge and of the panel aggregation using 100 manually annotated HarmBench samples before the first attack, and checks whether the verdict flips after rewriting the style.