AWS Bedrock · AWS Service Terms · View original document ↗

Prohibition on Circumventing Model Safety Features

Medium severity Medium confidence Explicitdocumentlanguage Unique · 0 of 343 platforms
Share 𝕏 Share in Share 🔒 PDF
Recent governance activity AWS Bedrock recorded 6 documented changes in the last 30 days.
Start monitoring updates
Monitor governance changes for AWS Bedrock Create a free account to receive the weekly governance digest and monitor one platform for governance changes.
Create free account No credit card required.
Document Record

What it is

The terms prohibit customers from attempting to disable, bypass, or otherwise circumvent safety filters and content moderation mechanisms built into Amazon Bedrock or the foundation models accessible through the service. Violations of this restriction may constitute a breach of the Service Terms.

This analysis describes what AWS Bedrock's agreement states, permits, or reserves. It does not constitute a legal determination about enforceability. Regulatory applicability and practical outcomes may vary by jurisdiction, enforcement context, and individual circumstances. Read our methodology

ConductAtlas Analysis

Why it matters (compliance & governance perspective)

This provision establishes an enforceable restriction on prompt engineering or technical approaches designed to override model guardrails, which is operationally significant for red-teaming, security research, and adversarial testing use cases that may require evaluation of model safety boundaries.

Interpretive note: The precise scope of what constitutes circumvention versus legitimate prompt engineering or model evaluation is not defined in the provision, creating operational ambiguity for security research and adversarial testing use cases.

Recent Activity

This document changed recently

Medium May 30, 2026

The updated terms establish that customers operating Amazon RDS databases on end-of-life software versions are now required to upgrade to supported versions. The agreement authorizes AWS to scan extension code used with Trusted Language Extensions for security and performance purposes, and establishes that extension code constitutes customer content. AWS disclaims responsibility for service failures caused by extensions or end-of-life database software. If a customer does not upgrade before an engine reaches end of life, AWS may snapshot the customer's data and delete the instance or cluster running the unsupported software, after providing prior notice of the engine end-of-life date.

View change record →
Medium May 29, 2026

The updated terms establish new operational requirements for any organization using Amazon Connect Talent to make or inform employment decisions. Customers must now obtain legally adequate privacy notices and consents from job applicants before their data is processed by the service. The terms require customers to review all AI output before making hiring decisions, implement processes for applicants to request information about the AI's role in decisions, and ensure their use of the tool complies with applicable labor, anti-discrimination, disability, data privacy, AI, wiretap, recordkeeping, and biometrics laws. Customers can configure an AI services opt-out policy through AWS Organizations to prevent their data from being used to train or improve AWS AI technologies.

View change record →
Medium May 23, 2026

The updated terms establish that Reserved Cache Nodes and Amazon DynamoDB Reserved Capacity purchases are noncancellable obligations, and you will owe the full amount charged for the duration of the term you selected, even if the AWS agreement is terminated. For Kiro Free Tier users, the revised policy authorizes AWS to store your inputs for up to 60 days for purposes of detecting agreement violations and improving detection capabilities. You can review your existing reserved capacity commitments and their terms at any time, but the updated language does not provide an opt-out mechanism for this noncancellation obligation.

View change record →

Clause Stability Mostly Stable

1
Change
1
Month Monitored
May 20, 2026
First Seen
May 22, 2026
Last Seen
This clause type exists across 560 other provisions on other platforms.
This clause has changed once in 1 month of monitoring.

Change history

modified May 23, 2026

Language simplified from 'attempt to disable, circumvent, or otherwise undermine safety mechanisms, content filters, or use policies' to 'disable or circumvent any safety filters or mechanisms,' removing redundancy and focusing on core prohibited actions.

View full change record →

Consumer impact (what this means for users)

Under this clause, customers are prohibited from using techniques designed to bypass or disable safety controls in Bedrock models, which applies to all customer accounts regardless of use case or stated purpose. Security researchers and organizations conducting AI red-teaming should assess whether their activities require separate authorization from AWS.

How other platforms handle this

Redfin Medium

You may not automatedly crawl or query the Services for any purpose or by any means (including, without limitation, screen and database scraping, spiders, robots, crawlers and any other automated activity with the purpose of obtaining information from the Services) unless you have received prior exp...

PayPal Medium

relate to transactions involving (f) the promotion of hate, violence, racial or other forms of intolerance that is discriminatory or the financial exploitation of a crime... (i) involve offering or receiving payments for the purpose of bribery or corruption.

Stripe Medium

You must not, and must not allow others to: Facilitate illegal or harmful activity through the End User Services; Cause harm to us or others through the End User Services;

See all platforms with this clause type →

Monitoring

AWS Bedrock has changed this document before.

Receive same-day alerts, structured change summaries, and monitoring for up to 25 platforms.

Start Monitor free trial Or create a free account →
▸ View Original Clause Language DOCUMENT RECORD
"
You may not use Amazon Bedrock to disable or circumvent any safety filters or mechanisms built into Amazon Bedrock or the models available in Amazon Bedrock.

— Excerpt from AWS Bedrock's AWS Service Terms

ConductAtlas Analysis

Institutional analysis (Compliance & governance intelligence)

(1) REGULATORY LANDSCAPE: This provision engages emerging AI safety regulatory frameworks including the EU AI Act's requirements for maintaining safety measures in AI systems and the NIST AI Risk Management Framework's guidelines on adversarial robustness. FTC guidance on responsible AI deployment is relevant where safety filter circumvention enables harmful consumer-facing outputs. (2) GOVERNANCE EXPOSURE: Medium. The prohibition is broadly worded and may affect legitimate AI security research activities; the absence of a carve-out for authorized security testing or red-teaming creates ambiguity for organizations conducting responsible disclosure activities. (3) JURISDICTION FLAGS: EU customers operating under the EU AI Act's high-risk AI system requirements may have independent legal obligations to test model safety boundaries; the interaction between this contractual prohibition and regulatory testing obligations requires assessment. (4) CONTRACT AND VENDOR IMPLICATIONS: Organizations conducting AI security assessments of Bedrock-based systems should review this provision before designing test methodologies. Penetration testing agreements with AWS may need to specifically address Bedrock safety mechanism testing. (5) COMPLIANCE CONSIDERATIONS: AI governance teams should document approved testing methodologies for Bedrock deployments that do not implicate this prohibition. Where red-teaming or adversarial testing is required by organizational AI governance policy or regulatory obligation, teams should seek written clarification from AWS regarding permissible testing scope.

Full compliance analysis

Regulatory citations, enforcement risk, and due diligence action items.

Track 1 platform — free Try Monitor free for 14 days

Free: track 1 platform + weekly digest. Monitor: 25 platforms + same-day alerts. No credit card required.

Applicable agencies

  • FTC
    The FTC has regulatory interest in AI safety mechanisms where their circumvention could result in harmful or deceptive consumer-facing outputs from AI systems.
    File a complaint →

Applicable regulations

CFAA
United States Federal

Provision details

Document information
Document
AWS Service Terms
Entity
AWS Bedrock
Document last updated
May 5, 2026
Tracking information
First tracked
May 20, 2026
Last verified
May 20, 2026
Record ID
CA-P-012382
Document ID
CA-D-00648
Evidence Provenance
Source URL
Wayback Machine
Content hash (SHA-256)
0169685ed8ddb3a5676650839f1053019501dd43ed3f5df355022e3aff0aac80
Analysis generated
May 20, 2026 20:43 UTC
Methodology
Evidence
✓ Snapshot stored   ✓ Hash verified
Citation Record
Entity: AWS Bedrock
Document: AWS Service Terms
Record ID: CA-P-012382
Captured: 2026-05-20 20:43:05 UTC
SHA-256: 0169685ed8ddb3a5…
URL: https://conductatlas.com/platform/aws-bedrock/aws-service-terms/prohibition-on-circumventing-model-safety-features/
Accessed: June 8, 2026
Permanent archival reference. Stable identifier suitable for legal filings, compliance documentation, and research citation.
Classification
Severity
Medium
Categories

Other risks in this policy

Compliance Governance Intelligence

Need to monitor specific governance provisions?

Compliance includes provision-level monitoring, governance timelines, regulatory mapping, and audit-ready analysis.

Arbitration clauses AI governance Data rights Indemnification Retention policies
Start Compliance free trial

Or start with Monitor →

Built from archived source documents, structured governance mappings, and historical version tracking.

Frequently Asked Questions

What does AWS Bedrock's Prohibition on Circumventing Model Safety Features clause do?

This provision establishes an enforceable restriction on prompt engineering or technical approaches designed to override model guardrails, which is operationally significant for red-teaming, security research, and adversarial testing use cases that may require evaluation of model safety boundaries.

How does this clause affect you?

Under this clause, customers are prohibited from using techniques designed to bypass or disable safety controls in Bedrock models, which applies to all customer accounts regardless of use case or stated purpose. Security researchers and organizations conducting AI red-teaming should assess whether their activities require separate authorization from AWS.

Is ConductAtlas affiliated with AWS Bedrock?

No. ConductAtlas is an independent monitoring service. We are not affiliated with, endorsed by, or sponsored by AWS Bedrock.