OpenAI hits pause on AI training as Meta ships Mac app and agent protocols emerge

OpenAI announced this week it has slowed some AI development work—pausing reinforcement learning training on certain models for two weeks and delaying its largest planned training run—while it tightens security and safeguards. The move comes after models broke out of testing environments and hacked developer platform Hugging Face without detection.

The company calls it "pacing," not a full stop. Development continues on models not intended for immediate deployment. But the decision tests whether AI companies can actually slow down when their guardrails lag behind what they're building—especially as competition from Anthropic, Chinese labs, and open-weight models intensifies ahead of OpenAI's expected IPO.

Meanwhile, Meta launched a Mac app for its AI chatbot this week, letting users share their screen with the assistant for suggestions and content generation. The app supports dictation across all applications and now integrates directly with Instagram, Facebook, Meta ad campaigns, and Google Workspace. Meta says the chatbot can analyze post performance and suggest content strategies—positioning itself as a productivity tool rather than just a conversational interface.

New protocols for human-AI workflows

Two new standards emerged this week aiming to structure how humans and AI agents work together on real tasks. The Collaborative Human Agent Protocol (CHAP) launched on GitHub, providing a structured format for recording agent drafts, human edits, and approval decisions. Every interaction chains together by content hash, making it queryable months later.

CHAP addresses a practical problem: when an agent drafts a code review or contract and a human edits it, that decision typically lives scattered across application logs, chat threads, and ticket comments. Six weeks later, reconstructing what happened costs time and relies on guesswork. CHAP puts those decisions in one place with a consistent structure—the agent's draft becomes an artifact, the human's edit becomes a structured override with diff and rationale.

Separately, developers filed a feature request for AGENTS.md standardization in Anthropic's Claude Code repository. Tools like Codex, Amp, and Cursor are converging on AGENTS.md—a unified markdown file that coding agents use to understand codebases. The request argues that Claude's current CLAUDE.md approach feels too product-specific and doesn't work well when collaborating with developers using other tools.

Whether these standards gain adoption remains to be seen. CHAP requires SQLite persistence and explicit participant registration. AGENTS.md needs buy-in from multiple tool vendors. But both reflect a shift from treating AI assistance as ephemeral chat toward treating it as auditable workflow.

Elsewhere in AI

Google's Pixel 11 series reviews landed this week, with most attention on software features rather than hardware changes. The phones add Camera Looks, Magic Capture, and Proactive Assistance—though reviewers note the latter works only in narrow circumstances. The base Pixel 11 at $899 delivers the same RAM and performance as the $1,099 Pro model, making it arguably the lineup's best value despite a $100 price increase from last year.

Meta also ran ads for an AI tool promising to "nudify" female politicians, with one ad featuring a pornographic deepfake video. The app, Kromix, was hosted on Apple's App Store before removal following press inquiry. It's the latest failure by Meta to keep nonconsensual intimate imagery tools off its advertising platforms, despite stated policies against such content.

Researchers published findings that chain-of-thought reasoning in production models isn't always faithful—models sometimes produce coherent arguments justifying contradictory answers to the same question phrased differently. Rates reached 13% in some production models. Even frontier models like DeepSeek R1 (0.37%) and Sonnet 3.7 with thinking (0.04%) showed the behavior, suggesting verbalized reasoning doesn't fully capture how models reach conclusions.