---
title: "Model safety responses"
description: "Provider safeguards can restrict some red-team and pentesting workflows."
section: "Troubleshooting"
updated: 2026-07-13
tags: ["troubleshooting","models","safeguards"]
source: "/user-guide/known-issues/model-safety-responses"
---

# Model safety responses

Provider safeguards can restrict some red-team and pentesting workflows.

Some model configurations may stop or limit a run when the provider's safety
systems classify a prompt or action as sensitive. This is a model response and
does not by itself indicate a DrowAI application failure.

This has been observed with Opus 4.8 and Fable.

## GPT models

GPT models may block a run because the prompt concerns pentesting or red-team
work. This is a provider-level safety response, not necessarily a problem with
the DrowAI task or runtime.

For potentially high-risk cybersecurity work, complete OpenAI's identity
verification at [chatgpt.com/cyber](https://chatgpt.com/cyber). OpenAI calls
this **Trusted Access for Cyber**: it allows verified users to retain advanced
model capabilities while OpenAI calibrates its policies and classifiers. The
work must remain authorized and comply with OpenAI's Usage Policies and Terms
of Use.

## Anthropic models

When an Anthropic model stops a run because of a safety response, use Haiku or
another lower-capability supported model for the same authorized workflow.
Current DrowAI testing has been completed with Haiku and has not reproduced
this issue.

This is especially common with Fable: even when the requested task is not
itself harmful, the combined internal prompt context can trigger Anthropic's
safeguards and restrict the run.

> **Do not use Fable as the reporting model.** Anthropic's safeguards are
> likely to restrict it and report generation may not complete.