OpenAI upgraded its developer toolkit on June 3, 2025, granting its Codex AI agent internet access within ChatGPT. This pivotal change expands Codex’s capabilities but also introduces new security considerations.
Codex is now available to ChatGPT Plus users with new features like pull request updates and voice dictation. OpenAI also enhanced its voice agent tools, including a new TypeScript SDK and an improved speech model.
Codex Connects: Expanded Capabilities and Security Trade-offs
Central to improvements is Codex’s new internet access. Previously absent when the agent initially launched in ChatGPT in May, this capability empowers it to install software packages, retrieve current documentation, and run tests requiring external data.
By default, internet access is disabled, though OpenAI provides users granular control over which domains and HTTP methods Codex can use, according to OpenAI. Such connectivity is now available for ChatGPT Plus, Pro, and Teams users, with Enterprise access to follow.
The decision to provide Codex with internet access involves “complex tradeoffs,” OpenAI CEO Sam Altman stated, urging users to “read about the risks carefully.”
codex gets access to the internet today! it is off by default and there are complex tradeoffs; people should read about the risks carefully and use when it makes sense.
also, we are making in available in the chatgpt plus tier.
— Sam Altman (@sama) June 3, 2025
OpenAI’s own documentation highlights security risks such as prompt injection and potential data exfiltration. Willison’s analysis further notes that while safeguards like domain allowlists are in place, the default “Common dependencies” list contains numerous domains. To bolster security, OpenAI advises restricting network requests to safer methods like GET, HEAD, and OPTIONS.
Broader Enhancements for Codex and Voice AI
Beyond internet access, OpenAI has widened Codex’s availability to ChatGPT Plus users, who will initially receive generous usage limits, subject to rate-limiting during high demand. Developers can now also direct Codex to update existing pull requests and can dictate tasks using voice input.
Further refinements, detailed in the OpenAI Codex Changelog include better support for binary files, improved error messages, an increased task diff limit to 5 MB, and a longer 10-minute duration for setup script execution.
OpenAI’s voice agent development tools also received substantial upgrades. The new Agents SDK, now in TypeScript, offers streamlined handoff mechanisms, configurable guardrails, detailed tracing for debugging, and critical support for human-in-the-loop approvals. The updates aim to give developers more power and flexibility.
Furthermore, an updated speech-to-speech model, accessible via new API endpoints (gpt-4o-realtime-preview-2025-06-03 and gpt-4o-audio-preview-2025-06-03), promises improved instruction-following, tool-calling consistency, and interruption behavior, along with customizable voice speeds. The Traces dashboard now also supports Realtime API sessions.
Strategic Vision and Developer Community Impact
The updates represent a significant evolution from Codex’s May 2025 debut. The agent initially operated always in an isolated cloud container, devoid of direct internet or external API access.
The introduction of controlled internet access directly addresses a key previous limitation. OpenAI’s broader strategy in the AI coding arena has been assertive, including a deal to acquire AI coding startup Windsurf and the April release of the separate open-source Codex CLI tool.
Eenabling developers to build robust and trustworthy AI systems is a top priority for OpenAI, and viewing the new tools, especially the human-in-the-loop features, are a critical steps in that direction.
OpenAI envisions these agents becoming “virtual teammates,” completing tasks autonomously that take human engineers hours or even days to accomplish. However, the company also stresses the ongoing need for human oversight, stating in its original Codex announcement that “it still remains essential for users to manually review and validate all agent-generated code before integration and execution.”
Looking ahead, Sam Altman has indicated that “future iterations will aim for ‘deep, proactive collaboration with development teams, not just reactive task completion,'” signaling further evolution in AI-assisted software development.
The OpenAI Team stated in their blog post their goal is to provide developers with advanced and flexible tools for responsible innovation.
