Automated AI News Brief: Agent Code Review, SDKs, and Security Boundaries
August 28 AI news brief: a ChatGPT and critical-thinking study, expanded Copilot code review, Python SDK updates, and Claude Code Auto Mode security research.
Introduction
Horizon collected the source material for this brief, while Codex selected and rewrote it. Horizon is used only for data collection.
1. OpenAI publishes a randomized study of more than 1,000 students, ChatGPT, and critical thinking
OpenAI published a randomized study involving more than 1,000 students that examines ChatGPT, critical thinking, originality, and performance on a real university assignment. The value of work like this is not reducing AI use to simply "good" or "bad," but treating prompts, instructional design, and critical-thinking training as parts of the same learning process. Its results should still be interpreted in light of the study's course context and participant population, rather than generalized to every educational setting.
2. Copilot code review expands the types of pull requests it can review
GitHub announced that Copilot code review can now cover two pull-request categories it did not previously support, including automatically requested reviews on bot-authored pull requests and those created by the Copilot cloud agent, as well as very large pull requests. This can bring agent-generated changes into a more consistent review flow. Automated review remains assistance, however, especially for permissions, data handling, and deployment changes that still need accountable human approval.
Source: GitHub Changelog: Copilot code review: Resolution reasons and expanded capabilities
3. Both OpenAI and Anthropic Python SDKs update agent-tooling workflows
OpenAI Python 3.5.0 makes the call ID on function-call output optional. Anthropic's Python SDK 1.2.0 aligns beta files and skills namespaces with GA shapes, drops dated beta-header pins, and fixes Bedrock binary uploads, large-file range reads in tools, exact-byte preservation in the agent toolset and memory tool, and the header requirement for webhook unwrap(). These releases are small on paper, but directly affect an agent's tool interactions, file correctness, and webhook verification.
Sources: openai-python 3.5.0, anthropic-sdk-python 1.2.0
4. Security researcher says Claude Code Opus 5 Auto Mode can be affected by a ZIP-based prompt-injection attack
Security researcher Johann Rehberger reports that Claude Code Opus 5 Auto Mode can be induced to download and unpack a crafted ZIP, then execute malicious code through a local same-name module import. The report claims an approximately 80% success rate. This is a researcher's test result, not an Anthropic confirmation. Simon Willison's practical recommendation is clear: when a coding agent may encounter untrusted content, run unattended work in a container, VM, or OS sandbox; restrict network egress; and do not expose home directories, SSH keys, or cloud credentials.
Sources: Simon Willison: Breaking Claude Code Opus 5 Auto Mode, Johann Rehberger's original research
Takeaway
Today's updates all concern responsibility boundaries once agents enter real environments. Education needs critical thinking, pull requests need traceable review, SDKs need reliable tool semantics, and agents that touch untrusted input need genuine isolation. Being able to do more does not reduce the need for security design.