> ## Documentation Index
> Fetch the complete documentation index at: https://docs.hiddenlayer.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Control Vector

## Detection Summary

The computational graph contains artifacts consistent with the insertion of a control vector into the model’s computational graph, which may alter or override expected model behavior.

## Security Impact

Control vectors can be used to modify model behavior at inference time, including suppressing refusal mechanisms or introducing attacker-defined responses, potentially enabling misuse, policy bypass, or unauthorized behavior in secured models.

## False Positive Considerations

Control vectors may be introduced intentionally for legitimate purposes such as fine-tuning, alignment adjustments, or experimental research, and are not inherently malicious without additional context.

## Recommended Remediation

* Treat the model as untrusted until the purpose and impact of the control vector are fully understood.
* Engage the team responsible for model development or sourcing to determine whether the control vector was intentionally introduced and for what purpose.
* Have personnel with appropriate ML and security expertise review the model architecture and parameters to assess how the control vector affects model behavior.
* Evaluate whether the control vector modifies refusal logic, safety constraints, or other guardrails in ways that could increase risk.
* If feasible, test the model in a controlled or sandboxed environment to observe behavioral differences with and without the control vector applied.
* If the control vector cannot be validated as legitimate or aligned with organizational policies, remove the model from the deployment pipeline.
* If the model is already deployed, escalate in accordance with established security and incident response procedures to determine appropriate containment, remediation, or rollback actions.