Secure network access for agents with strict allowlists
Combine skills with open network access cautiously; use strict network allowlists and domain secrets.
If you use networking, keep network allowlists strict, assume tool output is untrusted, and avoid open internet plus powerful procedures in consumer-facing flows where users expect strong confirmation controls. For hosted shell workflows, treat /mnt/data as the standard place to write outputs you’ll retrieve, review, or pass back into subsequent steps. Examples include reports, cleaned datasets, and finished spreadsheets. A good mental model: tools write to disk, models reason over disk, developers retrieve from disk. If a request includes domains outside the org allowlist, it will error. If an allowed domain needs auth headers, use domain_secrets so the model never sees raw credentials. At runtime, the model sees placeholders (for example, $API_KEY), and a sidecar injects the real values only for approved destinations. This is a strong default any time your agent needs to call a protected API from within a container.
Combining skills with open network access creates a high-risk path for data exfiltration.
Strictly controlling network access and handling credentials securely prevents unauthorized data exfiltration and protects sensitive APIs. Using `/mnt/data` for outputs provides a clear, auditable location for agent-generated files.
Prevent data exfiltration by implementing strict network allowlists when agents use networking capabilities. Use domain secrets for authentication and treat `/mnt/data` as the standard output directory for retrieved, reviewed, or passed-back artifacts.
Restricting network access to only approved domains and securely managing credentials via placeholders prevents malicious data transfer. Designating a standard output directory simplifies artifact management and review.
- Shell + Skills + Compaction: Tips for long-running agents that do real work ↗OpenAI Developers Blog · OpenAI Developers Blog · Sep 3, 2026 · quoted as shared
First seen on Subvolts Sep 3, 2026. Updated Sep 11, 2026. Spotted a problem? Tell us.
- Shell + Skills + Compaction: Tips for long-running agents that do real workOpenAI · OpenAI Developers Blog · Sep 3, 2026
The Playbook: reusable how-tos, prompts, configs and tips pulled out of the news.
- TemplateEvals CSV prompt list for agent skillsA CSV file template for listing prompts used in evaluating agent skills. · 20 min
- ChecklistAX Practitioner Playbook Evaluation Quality ChecklistReview your AI agent evaluation process before running it to ensure its quality and trustworthiness. · 30 min
- Skill or agent configJoplin agent for automated note filing and retrievalAutomate note organization and retrieval in Joplin using a custom agent and Python scripts. · 60 min
- HeuristicUse reasoning models for improved reliability on complex tasksEmploy reasoning models like o1, o3, and o4-mini for dramatically improved reliability on complex, multi-step work. · 5 min
- Setting tipStore AI agent traces and evaluation data with projectKeep traces, evaluation datasets, and results with your project for repeatable testing. · 5 min
- WorkflowBuild and test an AI agent using Microsoft Foundry ToolkitCreate, configure, test, and evaluate an AI agent within VS Code. · 60 min
Questions people ask
- Which AI tools does this work with?
- Any current model. It was seen with in the sources.
- Where did this come from?
- OpenAI Developers Blog (OpenAI Developers Blog). Subvolts collected, normalized and credited it; the link above opens the original.