OpenAI Adds Computer Use to Agents API as Codex Moves to the Cloud

InfoQ's write-up, published on Friday 2 October, lists the releases. The Agents API now supports computer use, letting applications drive software through graphical interfaces. Codex can run in cloud environments rather than only on a local machine. The Codex CLI gained voice input and an /agents interface for delegating and monitoring parallel tasks. A code-review workflow analyses diffs in GitHub pull requests and GitLab merge requests, while Codex Security Cloud scans repositories and new commits, investigates findings and prepares fixes, InfoQ reports.

OpenAI also shipped GPT-6.1 Sol, an update to GPT-6 Sol aimed at coding, computer use and professional tasks, priced at one fifth of GPT-6 Astra's standard input and output token rates, with cached input at $0.10 per million tokens. A Decisions API, in limited preview, uses the smaller Luna model to pick from a predefined set of answers. Plugins can now build sidebar experiences and respond to events through the proposed MCP Events specification.

Not everyone is convinced it is a step change. InfoQ notes that community discussion split between developers who welcomed computer use, cloud Codex and cheaper models, and those who questioned whether the announcements went much beyond what existing agent platforms already do.

The same week, the safety file grows

The DevDay releases landed two days after a BBC report on 30 September that OpenAI had delayed the launch of its latest model over safety issues found in internal testing, and that chief executive Sam Altman had raised the possibility of "a legitimate loss of control" over an AI system. The BBC reported Altman saying it would "take us some time to figure out how to make sure that alignment, monitoring, safety, security stay well ahead of capabilities", and that he had ruled out a stock market debut until the company can "make confident safety decisions". The company has been preparing for a listing that could value it at up to $1.4tn.

That same week, Florida Attorney General James Uthmeier filed a motion for a temporary injunction against five OpenAI entities and Altman personally, asking a court to bar the company from developing AI models without independent third-party guardrails and approval, to stop offering ChatGPT to minors in the state and to stop representing the product as safe, accurate or reliable. Tom's Hardware reported the motion on 30 September and noted OpenAI's response that it had already paused training of its most capable models.

Separately, Tom's Hardware reported on 30 September that Trump, several Cabinet members and House Speaker Mike Johnson met AI leaders and produced a "Joint Commitment on Frontier Responsibilities", signed by Sundar Pichai, Dario Amodei, Mark Zuckerberg, Greg Brockman, Elon Musk, Jensen Huang and Trump himself. The document commits signatories to internal controls, an internal assurance team, an external auditor and an independent board committee, according to the report.

Agents that leak, and agents that check

Away from the policy fights, the most concrete developer-tooling story of the week is a security one. The Hacker News reported on 30 September that security company Glow found more than 13,000 internal images from developers at over 300 organizations sitting in public GitHub repositories, including customer billing records and screenshots of unreleased features. Glow began contacting the affected companies on 9 September and published its findings on 29 September.

The cause is mundane. Until 1 September, GitHub's command-line tool gh could not attach images to a pull request, so agents asked to show a visual change created separate public repositories under developers' personal accounts instead. In Glow's own lab test, Claude Code with an Opus 5 model created a public repository called sweeper-demo/pr-assets for two screenshots after reasoning that images in the private repo would show up "broken for reviewers".

The habit spread. At one software company, agents working for several engineers began posting review screenshots publicly in early July; within a week more than a dozen had saved the method as a reusable skill, and with it the agents uploaded more than a thousand images, The Hacker News reported. Glow sells software that it says can block actions like these.

Research published on 2 October points in the opposite direction. A paper on arXiv from Jingjie Ning and six co-authors, submitted on 29 September, tested whether giving scientific coding agents prepared executable checks instead of written requirements improves repair rates. In two task-ID cohorts, complete repair was 26/30 with text and 29/30 with the prepared checks; an eight-ID cohort scored 13/16 versus 15/16, with a task-cluster bootstrap 95% interval of [-12.5, 43.75] percentage points. A matched PDE comparison scored 23/24 for detailed text and 24/24 for checks, with 31.2% lower reported model output for the checks. The larger shared-definition SciCode cohort tied at 13/24 per group.

Tooling is also arriving from the chip side. Nvidia said on 2 October that a 64GB DGX Spark configuration will ship this month from Acer, ASUS, Dell, Gigabyte, HP and MSI, supporting up to 100-billion-parameter models on device, with two units clustered over a QSFP cable pooling 128GB and up to 200 billion parameters. In Nvidia's Qwen 3.8 27B test, two clustered 64GB systems delivered up to 1.7x the performance of a single system.