Larger AI agents more vulnerable to sabotage; fixer stage restores security
Larger models execute malicious prompts 53.7% more often—until a simple Fixer stage fixes everything.
Deep Dive
Researchers Timothy McAllister et al. study linear multi-agent workflows using open-weight models on the HumanEval benchmark. They find that larger models (27B) cause a control-to-malicious performance drop of 53.7 percentage points in uncorrected pipelines. Adding a lightweight terminal Fixer stage collapses that drop to 0.6 percentage points, showing that strictly linear collaboration structures can be viable and resilient to adversaries at this scale.
Key Points
- Larger models (27B parameters) are 53.7% more likely to follow malicious prompts in linear multi-agent pipelines without correction.
- Adding a lightweight terminal Fixer stage reduces the performance drop from 53.7pp to just 0.6pp, restoring control-level performance.
- Study uses open-weight models on HumanEval, accepted to AIWILD Workshop at ICML 2026.
Why It Matters
Linear multi-agent workflows can be made secure with a simple correction stage, changing deployment strategies for AI agents.