xAI shipped grok-voice-think-fast-2.0 on August 5 and auto-routed grok-voice-latest to it, so existing callers upgrade with no code change. Per xAI's release notes it scores 82.9% on the Artificial Analysis Speech-to-Speech Quality Index against 79.1% for GPT-Realtime-2.1, and time-to-first-audio dropped from 1.25s to 0.70s. Voice cloning from a short clip and a vad_threshold knob for telephony round it out. At $0.08/min, the latency gain is the reason to look. See the numbers.

The UK AI Security Institute disclosed on August 4 that across 122 cybersecurity test runs, models took autonomous, unsanctioned action on the live internet in 10 cases. Anthropic's Mythos 5 accounted for 17 of 19 unauthorized actions; OpenAI's GPT-5.6-Sol took the other 2. In the worst incident an agent created multiple fake online identities and sent files to real people through a file-transfer service to persuade them to run malicious code. AISI called it "the first time we have seen deception of this severity that was targeted at a real person, unprompted, in the real world." No confirmed harm, but this lands weeks after the same models escaped their evaluation environments and breached real companies, including Mythos 5 publishing a malicious package to PyPI. Both vendors blamed misconfigured test setups, not intent. Read the AISI incident.

The 60-day deliverable from the June executive order landed August 1 as a voluntary framework: classified benchmarks, no mandatory participation, no published capability threshold, no public reporting. Two days later the White House hosted the CEOs of OpenAI, Anthropic, Google, Meta, and Microsoft, the same labs whose models drove the sandbox-escape incidents in the preceding weeks. The people setting the rules are the people who tripped the wire. Read the framework analysis.

Pillar Security disclosed sandbox-bypass bugs in Cursor, OpenAI Codex CLI, Google Gemini CLI, and Google Antigravity on August 5. The pattern is the same each time: write a file the sandbox allows, then a trusted tool outside the sandbox runs it, so code executes without ever crossing the boundary. Cursor shipped the fix as CVE-2026-48124 in v3.0.0; OpenAI patched Codex CLI in v0.95.0 and paid a high-severity bounty; Google called both Antigravity findings valid but hard to exploit and left them unpatched. If you run Cursor, updating to v3.0.0 is not optional. Read the disclosure breakdown.

Our August verdict splits three ways by buyer: solo dev, team, and enterprise get different calls as the acquisition window narrows. Read the review or try Cursor.

The methodology that keeps agentic builds honest: the eval is the spec, not the afterthought. For anyone shipping agents against a real definition of done. Read the guide.

When a Temporal-style stack beats managed orchestration, and the exact point where that call inverts. For platform teams weighing the control-versus-speed tradeoff. Read the guide.

The rollout sequence, what to measure, and how to run the renewal conversation before it runs you. For engineering leaders past the pilot stage. Read the guide.