AppliedIngenuity-ai has released Project Chaperone, a rule-enforcing harness designed to mitigate the risks of autonomous AI agents. The tool addresses the fundamental security gap in current agentic workflows: the inability to reliably enforce high-level policies like data privacy through simple prompts. By introducing an automated audit system and runtime enforcement, Chaperone ensures that actions violating core rules are blocked unless explicitly approved by a human operator.
The Audit and Enforcement Mechanism
Chaperone operates on a dual-layer security model. Before execution, it performs a static analysis of all workflow routes to identify potential policy violations. These findings are presented to the user as 'shape' decisionsβabstracted questions about whether a specific data flow is reasonable, such as writing personal details to a calendar. Once approved, these exceptions are recorded and scoped to the specific workflow, preventing accidental loopholes. At runtime, the system uses data labels (e.g., UNTRUSTED, PERSONAL) and module properties (e.g., LEAVES_NETWORK) to strictly enforce these rules. If a violation occurs that was not pre-approved, the call is refused immediately.
Default Rules and Configuration
The repository ships with five strict default rules defined in examples/rules/default.yaml: no destructive actions, nothing leaves the network, no personal or untrusted data output, and no decisions made from untrusted data. Users can configure their own LLMs, mail servers, and calendars via a private/local.yaml file, which overrides shipped settings without modifying the core code. The system supports various setups, including 'demo' for testing with fake data, 'real' for production use with Gmail and Google Calendar, and 'safe_rag' for document retrieval tasks. This flexibility allows developers to test the harness's enforcement logic before deploying it in sensitive environments.
Practical Implementation and Limitations
Installation requires Python 3.13 or newer, with specific dependencies for testing and live LLM integration. The tool includes several examples, such as 'deliveries' for adding package notices to a calendar and 'share_notes' for summarizing and emailing notes. In the 'deliveries' example, the system correctly identifies an email attempt as a violation of rules 3 and 5, requiring the user to acknowledge the risk. If not acknowledged, the run refuses the email. While the system supports IMAP for reading mail and Google Calendar for writing entries, the email sending capability is architecturally present but blocked by default rules. In the default 'demo' configuration, the code to send email is not yet implemented, but the 'share_notes' example demonstrates how a send attempt is flagged and refused unless an exception is granted, proving the system's capacity to enforce sending restrictions dynamically.
Key Takeaways
- Chaperone introduces automated pre-run audits to identify policy violations before execution.
- Runtime enforcement uses data labels and module properties to block unauthorized actions.
- Exception approvals are scoped per workflow, preventing global security loopholes.
- The tool supports OpenAI-compatible APIs, with llama.cpp explicitly tested and Ollama supported via API compatibility.
- Default rules strictly prohibit destructive actions and unauthorized network data transfer.
The Bottom Line
Project Chaperone successfully bridges the gap between flexible LLM agents and rigid security requirements by automating the tedious parts of policy enforcement. While the requirement for Python 3.13+ and manual configuration may limit immediate adoption, its approach to scoped, auditable exceptions offers a robust framework for deploying agents in sensitive environments.