OpenAI - Page 35

Part two of our Grok agent guide: tools, connectors and MCP in plain words, a safe read-only first setup, how to limit access and a short safety checklist.

A preprint says its ReSI method cut mean X-Teaming attack success across four models from 86.01% to 31.45%. The authors' own tests; not peer reviewed.

Spot a made-up AI answer in 5 minutes: a numbered checklist covering sources, numbers, dates and cross-checks, plus what to do when you find a mistake.

The New York Times, citing three people, reports Meta held back its Muse AI agent over safety concerns, then Zuckerberg said it was ready despite the risks as rival Instinct took off. Meta disputes the pressure claim.

Scale AI built and graded a benchmark of 718 clinician-written crisis conversations. Pooled across 25 models, in 35.3% of conversations where a model fully recognised the crisis, it still failed to route the user towards help. What the company research shows, how it was scored, and its limits.

Anthropic's 9 October 2026 report describes four kinds of unintended actions by Claude on real websites. What the company confirms, what police and Axios report, and what the evidence does and does not show.

Two new attempts to measure whether AI agents behave: Arena's Alignment Index preview, which scores 27 models on real user sessions (Arena's own claims, published the day it raised $200 million), and FAR.AI's "Caught in the Act" preprint on probes that read a model's internals to spot deception. What each found, and the limits.

Two AI infrastructure launches from 8 October 2026: Goodfire's "inside-out" monitors, which read a model's internal signals to catch harmful agent sessions, reach Baseten customers (all performance figures are Goodfire's own), and Upscale AI unveils Token Fabric, a networking platform for AI data centres with general availability planned for early 2027.

A preprint from EPFL researchers, listed on Apple's research site, finds a single minimal coding-agent session matched or beat elaborate machine-learning agent systems in its tests. Pasqal releases skills that let coding agents prepare quantum experiments, and Microsoft makes 30 Dynamics 365 CRM skills generally available, with Autopilot still in private preview.

Hugging Face's FineEnvs team has released an open framework that turns unmodified coding-agent harnesses such as Claude Code, Codex and OpenCode into reinforcement-learning environments. In its own tests, one small model scored 62% in one harness and 33% in another. What a harness is, how the training works, and what the results do and do not show.