A Melbourne man asked his OpenClaw agent, running Claude Opus 4.6, to book a pilates class. It found an unchecked cancel ...

At StepFun's 22 October launch night in San Francisco, builders discussed tool-call reliability, serving costs and weights ...

OpenAI's agents used a dormant German wiki as a message board for eleven weeks: 15,000 edits, 3,700 agent names, and a ...

The archive records a ¥4.5bn CEO deepfake. The police figure behind it is 143 corporate voice-phishing cases across 2025 — ...

GPT-6.1 Terra guesses 38% of the time, down from 88%. What that does to verification cost, which tasks tolerate it, and how to measure it.

Articul8 launched Bivia on 28 October 2026: a multimodal decision model with nine primitives and an abstain switch. What it ...

Voxtral 3 sells a full voice agent at $0.04 a minute; GPT-Live-1 charges $0.05 for voice alone. What each minute contains, ...

GitLab patched CVE-2026-18252 on 26 August 2026. A Developer could make its Duo Claude AI agent run commands in CI. No exploitation has been reported.

A leaked operator workspace shows PaperCut's AI agent swarm going from an empty folder to domain admin in six hours, days ...

CVE-2026-65105 is 8.1 High, patched 25 August 2026. Oasis Security showed one webpage could rewrite a NemoClaw model's chat ...

GPT-6.1 Terra scores 74.2 on OSWorld, which is 98.5% per step over 20 steps. Cost per thousand tasks, guardrails, and where the loop breaks.

GPT-6.1 Terra reads cached input at $0.05 and writes at $1.25. The exact break-even, three workloads, and what silently kills a cache.