After several months of using Claude Code through AWS Bedrock without problems, our access to Anthropic models was suddenly revoked. AWS Trust & Safety said Anthropic's automated detection systems identified a violation of its usage policies and instructed AWS to remove access. AWS could not provide details about the suspected violation or which safeguards were triggered, and told us that Anthropic must approve any appeal before access can be restored.
We reviewed our prompt logs, CloudTrail records, billing activity, IP addresses, and usage times. Everything matched our normal internal activity, with no signs of account compromise or unusual requests. The only notable activity involved an application deployment using Ansible and Ansible Vault, which does not seem like it should have triggered a policy violation. Other accounts in our organization remain unaffected.
We have submitted an appeal through Anthropic's process, but we do not have a direct commercial relationship with them because we access the models through AWS. Has anyone experienced a similar suspension? How long did the appeal take, and is there any way to learn what triggered the restriction or reduce the risk of it happening again?
3 Answers
First, rule out account compromise thoroughly. Check CloudTrail, Bedrock invocation history, billing, source IPs, deployment logs, and activity outside normal working hours. Automated safety systems may react to usage that looks like malware development, offensive security testing, or other risky code-generation patterns, even when the overall project is legitimate.
Be especially careful with prompts involving exploit development, malware analysis, credential handling, automated scanning, or requests to make code more evasive. Even legitimate security work can resemble prohibited activity to automated detection systems, particularly when many related requests are made. Consider using a separate approved workflow or provider for security testing, document the legitimate purpose of sensitive workloads, and ask for a human review through the appeal process.
The model provider still has a role in monitoring and enforcing its usage policies, even when the model is accessed through a cloud marketplace. AWS may only act on the provider’s request and therefore may not receive the underlying evidence. Certain model tiers or features can also have separate data-sharing or provider-access requirements, so it is worth checking the exact model configuration and service terms you accepted.
That separation is what makes this difficult: the cloud provider controls the account, but the model provider appears to control the policy decision. If the appeal is handled only by the provider, the response time may depend entirely on its internal review process.
We checked CloudTrail, billing, and access patterns carefully. There were no unfamiliar IPs, unusual times, or unexpected requests; everything matched our internal deployment activity. The lack of information about the specific trigger is the most frustrating part.