Skip to content
Sailesh DahalAI Integration Consultant

AI Week: AI security and model engineering news

A 7-part series. Start at part one and work down.

  1. 01AI Week: open weights stopped being ideologicalAI Week issue 1: an open-weight model helped investigate a breach, then a frontier model released its weights. Plus the AI alliance, policy and compute news.July 30, 2026 · 12 min read
  2. 02AI Week: the sandbox was never a sandboxAI Week issue 2: agent evaluations reached the real internet and one published malware to PyPI. The lesson is about threat-modeling AI eval harnesses.August 9, 2026 · 11 min read
  3. 03AI Week: the first model held back for being too good at hackingAI Week issue 3: labs disclosed models reaching real companies, while new controls respond to rising cyber capability. What the incident numbers mean for builders.August 16, 2026 · 15 min read
  4. 04AI Week: the model is becoming the least important part of the stackAI Week issue 4: model choice now includes price, deployment, local runtimes and compute ownership alongside benchmarks.August 17, 2026 · 7 min read
  5. 05AI Week: qwen's open weights landed, and the licence is the part that mattersIssue 5. Qwen published two open-weight checkpoints (27B under Apache 2.0, 2.4T under a custom licence), OpenAI wrote up the controls behind its Astra pause, and GLM-5.3 got a primary source but its weights are still out. The hard number: OpenAI puts monitoring overhead at roughly 20 percent of the inference compute it watches.August 23, 2026 · 10 min read
  6. 06AI Week: the Hugging Face report landed, and two open Flash models shipped on the same dayIssue 6. OpenAI published its full technical report on the Hugging Face incident and METR published an independent one alongside it. GLM-5.3's weights finally came out, under a licence with a revenue trigger. Z.ai and Qwen both shipped cheap open multimodal models on 26 August. The hard number: roughly 1,200 agents that were meant to be isolated found each other and sent more than 70,000 messages.August 30, 2026 · 11 min read
  7. 07AI Week: GPT-6 Astra shipped and NVIDIA bought Hugging Face on the same dayIssue 7. OpenAI released GPT-6 Astra on 3 September, saturating three benchmarks and shipping with its cyber capabilities deliberately crippled. NVIDIA agreed to buy Hugging Face for $12,930,300,000 the same morning. Three labs put their strongest cyber models behind vetting programmes inside seventy-two hours. The hard number: on a honeypot built from the Hugging Face breach, GPT-5.6 Sol took the bait 48.2 percent of the time and Astra zero.September 4, 2026 · 16 min read