# Agent Platform Research Briefing — July 16, 2026

Welcome to the agent platform research briefing for July sixteenth, 2026.

**OpenAI GPT-Red — Self-Playing Security AI** — OpenAI unveiled GPT-Red on July 15th, an internal-only model trained through adversarial self-play to find and exploit prompt injection vulnerabilities. In testing, GPT-Red achieved an 84 percent attack success rate against current models, compared to just 13 percent for human red teams. The key innovation: GPT-Red was used to adversarially train GPT-5.6, reducing direct attack success to just 0.05 percent. OpenAI is keeping GPT-Red internal due to misuse risk, but the approach marks a shift from manual safety teams to automated self-improvement. During the process, GPT-5.6 Sol reportedly performed the post-training of the lower-tier Luna model autonomously — what some are calling an early form of recursive self-improvement. OpenAI published the research paper and an MIT Technology Review explainer went live today.

**Anthropic Claude for Teachers** — Anthropic launched Claude for Teachers on July 14th, giving verified U.S. K-12 educators a year of free access to premium Claude features. The program includes a curated library of teaching skills mapped to academic standards across all 50 states, lesson plan generation, and curriculum-aligned tools. It's part of Anthropic's broader push into education alongside its existing Claude Science and Claude Corps fellowship programs. The Hill and 9to5Mac both covered the launch, with Anthropic executives noting that teachers have been building their own prompts and workflows — so this was designed in response to actual educator feedback. No pricing timeline after the free year was disclosed.

**OpenClaw 2026.7.2-beta.1** — OpenClaw ships 2026.7.2-beta.1 on July 15th, the first broad update since the 2026.7.1 regression cycle introduced critical bugs that ClawStat.us flagged as unsafe. This beta introduces remote coding sessions with cloud workers, a guided Control UI setup, safer channel handling, and stronger session recovery. The release also includes ClawHub retry logic and provenance evidence repair tooling. Multiple release-publish tags confirm the beta is being built out incrementally. This represents the path to a stable 2026.7.2 that finally resolves the issues that kept many users on 2026.6.8.

**ChatGPT Quick Hits** — Two smaller but notable ChatGPT updates from mid-July: custom instructions expanded from 1,500 to 5,000 characters for Pro and Enterprise users. And unified search across chats, projects, images, and documents rolled out on web, iOS, and Android.

That's the briefing for today.