BRANCH paper bypasses multi-scanner AI guardrails
An arXiv paper posted 7 Oct 2026 describes BRANCH, a search method aimed at guardrails built from several scanners. It reports a 100% attack success rate across six guardrail systems in 120 scenarios, and transfer to 29 unseen guardrails, including eight commercial black-box products. The paper reports no fix.
Facts read on .
On 7 Oct 2026, researchers posted "BRANCH: Bypassing Multi-Scanner AI Guardrails" to arXiv. Multi-scanner guardrails combine several detectors, on the premise that shared representations across classification boundaries make known bypass techniques ineffective.
The authors propose a branching tree search that applies adversarial perturbation against individual scanners and chooses techniques by improvement across all scanners. They report a 100% attack success rate across 6 guardrail systems in 120 scenarios, with 72% fewer queries and 4.5x less wall-clock time than established techniques, while preserving the meaning of the text.
The paper also reports that bypasses made against its test systems transferred to 29 unseen guardrails, including 8 commercial black-box guardrails, in some cases raising success up to 100% with no extra optimization. These are test-setting results from the authors. The practical reading for buyers is to ask how a tool behaves when the text filter misses, and whether a second control gates the resulting action.
Sources
- arXiv 2610.10742: BRANCH, Bypassing Multi-Scanner AI Guardrails (arxiv.org), read