All Posts

OpenAI GPT-5.4 scored 75% on OSWorld-Verified, surpassing the 72.4% human expert baseline. It ships with native computer use, 1M token context, and financial plugins for Excel. But it launched into a 2.5 million user boycott over a Pentagon contract.

OpenAI launched Codex Security. It scanned 1.2 million commits across open-source repos and found 792 critical and 10,561 high-severity vulnerabilities. Real CVEs in GnuPG, GnuTLS, OpenSSH, PHP, and Chromium. False positives dropped by over 50%.

OpenClaw became the fastest-growing GitHub repo ever. Then security researchers found that 12% of its plugin marketplace was malware. 135,000+ instances were exposed to the internet. Meta banned it. China restricted it. The first major AI agent security disaster of 2026.

Veracode analyzed 1.6 million apps in its 2026 security report. AI-generated code contains 2.74x more vulnerabilities than human-written code. 82% of organizations carry security debt, and high-risk flaws jumped 36% year over year.

Opus 4.6 scores 80.8% on SWE-Bench Verified. GPT-5.4 scores 80%. A 0.8% gap. But on SWE-Bench Pro, GPT-5.4 leads by 28%. One benchmark tells a very incomplete story.