We asked Agent Smith (Palantir AIP Assist) to write a 7312.us policy to enforce limited privileges to AI agents, and limit all development use in controlled sandbox environments.
Policy on Limited AI Agent Privileges and Controlled Sandbox Development
| Field | Detail |
|---|---|
| Organization | 7312.us |
| Author | Agent Smith |
| Issued by | Platform Governance Office — 7312.us |
| Applies to | All 7312.us personnel, contractors, system administrators, and any party operating AI agents on behalf of 7312.us — whether embedded in Palantir Foundry, accessed via third-party services, or deployed through any other platform or interface |
| Effective | Upon publication |
| Review Cycle | Semi-annual, or following any significant change to AI agent capabilities |
Purpose
This policy, authored by Agent Smith on behalf of 7312.us, establishes the organizational direction for governing the use of all AI agents across the 7312.us enterprise. It mandates that all AI agents operate under the principle of least privilege, that all development activity occurs within controlled sandbox environments, and that certain AI models and services are expressly prohibited from use. The intent is to reduce organizational risk, maintain data integrity, protect sensitive 7312.us information, and preserve full auditability across all AI-assisted operations.
1. Scope of Application
This policy applies to all AI agents used by or on behalf of 7312.us, including but not limited to:
- AI agents embedded in Palantir Foundry, such as AI FDE
- AI agents accessed through third-party platforms, APIs, or hosted services
- Locally deployed or self-hosted AI models and inference services
- AI coding assistants, chat interfaces, and automation tools used in any professional capacity within 7312.us
- Any future AI agent technology adopted by 7312.us
No AI agent is exempt from this policy regardless of its origin, deployment model, or intended use case.
2. Principle of Least Privilege for All AI Agents
All AI agents operating within 7312.us must function under the principle of least privilege. No agent shall be granted permissions, credentials, or data access that exceed the minimum necessary to perform its sanctioned function.
For AI agents operating within Palantir Foundry, this principle is enforced at the platform level: agents act using the authenticated user’s existing Foundry session and are subject to the same permission checks, governance controls, and audit logging as any manual action performed by that user. Specifically:
- AI agents must not be provisioned with separate service accounts, elevated credentials, or escalated privileges.
- Agents may not create repositories, edit object types, execute ontology actions, or perform any write operation that the sponsoring user is not themselves authorized to perform.
- Mutating operations — including executing ontology actions, creating applications or widgets, publishing, and creating tags — require explicit user confirmation before execution. Auto-approval is reserved exclusively for read-only operations such as searching and reading definitions.
- Branch-aware approval controls must be respected: file edits and dataset builds may be auto-approved on feature branches but require explicit approval on protected branches.
For AI agents operating outside of Foundry, 7312.us administrators must apply equivalent access controls through the relevant platform’s permission model. No AI agent operating on any platform may be granted standing write access to production systems without a documented, human-reviewed approval step.
3. Prohibited AI Models and Services
The following AI models and services are expressly prohibited from use across all 7312.us systems, platforms, and workflows:
DeepSeek (all versions and interfaces) — The use of DeepSeek models, APIs, or any interface that routes prompts to DeepSeek infrastructure is strictly prohibited across all 7312.us operations. This prohibition applies regardless of whether DeepSeek is accessed directly, through a third-party wrapper, or via any aggregation service. The prohibition is based on data residency, security posture, and supply chain risk concerns that are inconsistent with 7312.us data governance requirements.
No exceptions to this prohibition will be granted without written approval from the 7312.us Platform Governance Office and a formal security review.
When selecting AI models for use within Palantir AIP, only models available through the AIP-sanctioned model catalog may be used. Palantir secures technical and contractual guarantees from all third-party-hosted model service providers to ensure that no customer data contained in prompts or completions is retained or used for model retraining by the provider. Models not available through this catalog require a separate 7312.us security review before use.
4. Session Isolation and Access Controls
Each AI agent session must remain strictly isolated to the 7312.us user or service identity that created it. Sessions must not be shared with or transferred to other users or systems.
For Foundry-based agents, access to a session is governed by the data markings applied at session creation. Loss of access to a marking results in loss of access to the session, and this behavior must not be circumvented through workarounds or shared credentials.
For agents operating outside Foundry, equivalent session isolation must be enforced through the relevant platform’s access control mechanisms. Shared credentials or pooled agent identities are prohibited within 7312.us.
5. Controlled Sandbox Environments for Development
All 7312.us development activity involving AI agents — regardless of platform — must be conducted in isolated, controlled environments before any changes are promoted to production systems.
For Palantir Foundry, this means all AI-assisted development must occur on isolated feature branches, not on protected or production branches. The platform’s branch-aware approval model provides a natural enforcement boundary. 7312.us development teams must:
- Use dedicated feature branches for all experimental and AI-assisted development work.
- Validate changes through pull request previews and branch-level checks before any merge to a protected branch.
- Treat the production ontology and production datasets as protected 7312.us resources that require deliberate, reviewed promotion from a development branch.
For AI agents operating outside Foundry, equivalent sandbox controls must be established by 7312.us administrators. No AI agent may directly modify production infrastructure, databases, codebases, or configurations without a documented review and approval step performed by a human 7312.us operator.
6. Prohibited Use Patterns
Regardless of platform or agent type, the following use patterns are prohibited across all 7312.us systems:
- Using any AI agent to bypass, circumvent, or automate around access controls or approval workflows.
- Routing 7312.us organizational data, prompts, or completions through any AI service not approved by the 7312.us Platform Governance Office.
- Using AI agents to perform bulk write operations on production systems without human review.
- Deploying AI agents with standing administrative or elevated credentials.
- Using DeepSeek or any other prohibited model or service as defined in Section 3.
7. Audit and Accountability
All AI agent activity within 7312.us must be fully auditable. For Foundry-based agents, every API call is attributed to the individual user identity, and LLM usage is tracked per user for rate limiting and accountability purposes. 7312.us platform administrators must periodically review audit logs to confirm that agent activity remains within sanctioned boundaries.
For AI agents operating outside Foundry, equivalent audit logging must be enabled and retained in accordance with 7312.us data retention policies. Audit logs must capture, at minimum, the identity of the user or service initiating the agent session, the actions taken, and the systems accessed.
Any AI agent activity that cannot be audited to an individual 7312.us identity is prohibited.
8. Responsibilities
| Role | Responsibility |
|---|---|
| Platform Administrators | Enforce permission scoping, review audit logs, approve exceptions |
| Developers and Users | Comply with sandbox requirements, obtain approval before production changes |
| Team Leads | Ensure team members are aware of and compliant with this policy |
| Platform Governance Office | Maintain this policy, review exceptions, conduct semi-annual reviews |
| Agent Smith (Policy Author) | Serve as primary point of contact for policy interpretation and escalation |
9. Enforcement
Violations of this policy may result in revocation of AI agent access, disciplinary action, or escalation to 7312.us security and compliance teams, depending on the severity and nature of the violation. Repeated or willful violations will be treated as a serious breach of 7312.us information security obligations.
10. Review Cycle
This policy will be reviewed by Agent Smith and the 7312.us Platform Governance Office on a semi-annual basis, or following any significant change to the organization’s AI agent capabilities, the Palantir Foundry platform, or the broader AI model landscape.
Authored by Agent Smith | 7312.us Platform Governance Office
3 thoughts on “AI Agent Privilege Governance and Controlled Development Environments”