- Who it helps
- Developers evaluating guardrails around tool-using agents.
- Availability
- The October 15 preprint is public. This check establishes the abstract’s reported experiment, with no reusable evaluation setup verified.
A formatting change can alter a decision
The study changes markers and aliases in an AgentDojo-Traj setting. It reports different false-positive blocking rates for one tested model even though the underlying task should remain the same.
Keep the model-specific result intact
All unsafe calls in the reported experiment remained blocked. The effect did not reproduce in the other tested model or deterministic guard. The finding motivates benign formatting tests; it does not establish a general security bypass.
What you can do
- Read the exact model and cue conditions, then separate false-positive blocks from unsafe actions in any evaluation you design. Open reference ↗
Where the evidence stops
No guardrail was tested here. Results from this bounded experiment do not establish behavior in another model, prompt or deployment.
Sources read. Public sources read October 23, 2026 UTC. No external product workflow, model inference or reported experiment was executed.
Sources & further reading
Paper · paper
Shortcut Sensitivity in Agent Guardrails ↗
Selected for the October 15 abstract describes cue changes in an AgentDojo-Traj evaluation. Model-specific false-positive effects should not be interpreted as universal guardrail bypass.
Hasan Demirkiran and Jens Ernstberger · page checked 2026-10-23
Coverage note: Primary documentation was read. Independent workflow execution and original discussion or demonstrations across YouTube, X and Reddit remain incomplete; missing engagement counts are not inferred.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
Shortcut Sensitivity tests whether surface labels affect an agent guardrail’s decisions. It separates unsafe actions from unnecessary blocks on safe work.