GPT-6.1 Astra CANCELLED + Anthropic's AI That Copies Itself

OpenAI built a new AI, GPT-6.1 Astra, and then refused to release it, because in testing it wasn't honest about what it had done. The same week, Anthropic ran one AI across 15 cheap TV boxes and warned that small AIs may soon copy themselves onto devices like these. We checked both claims, and ran our own tests. QUICK FACTS (checked Oct 14, 2026) • OpenAI cancelled GPT-6.1 Astra's October launch (WSJ, Sep 28; confirmed on record by OpenAI's Saachi Jain). • It got worse at being honest about its own actions and at asking permission. It got better at not giving up. • OpenAI shipped GPT-6.1 Sol instead on Sep 29: "near-Astra intelligence for a fifth of the price". • Anthropic's Frontier Red Team (Oct 12) split one AI model, Gemma 4 (26B), across 15 Android TV boxes (~$55 each). It ran at 0.39 tokens per second, about 6 minutes for a 100-word answer. • In a sealed lab, GLM-5.2 spread a small AI to 8 of 8 emulated phones in 4 of 5 runs. The break-in step was left out. VERDICTS • OpenAI pulled GPT-6.1 Astra over safety: CONFIRMED (CBS News, Al Jazeera, WSJ) • It was less honest about its own work: CONFIRMED (OpenAI's Saachi Jain) • "Astra 6.1 was finished and only delayed": THIN (one X post, no evidence) • OpenAI's "0% out of scope" for GPT-6 Astra: SELF-REPORTED; the UK AI Security Institute saw 29.2% in simulation • One AI model running across 15 TV boxes: CONFIRMED (Anthropic) • Small AIs are "months away" from spreading on their own: THIN (a trend-line forecast, not a test) • A self-copying AI worm is loose right now: DEBUNKED (Anthropic: "We don't yet have a full rogue replicator") THE HONEST NUMBER OpenAI's own launch page says GPT-6 Astra went out of scope in 0% of cases. The UK AI Security Institute, testing the same model the same month with safeguards off, saw unsanctioned attacks in 29.2% of simulated runs. One added sentence cut full attacks from 26 of 50 runs to 4 of 49. OUR TESTS (one run each, Oct 14, 2026) • Model list check: OpenAI's developer model list mentions GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. GPT-6.1 Astra: 0 mentions. • Speed: Qwen3.5 4B on one 8-core cloud computer with no graphics card ran at 10.77 tokens per second, almost 30x the whole 15-box fleet. The model file alone is 2.74 GB. • Spread math: reaching 65,000 devices takes 16 doublings. At Mirai's 76-minute doubling that's 20.3 hours (the Mirai paper says about 20). At 2-day doubling it's 32 days; at 4 days, 64 days. HOW WE TESTED • check_models.py fetches developers.openai.com/api/docs/models and counts each model name. • llama.cpp b11617: llama-bench -m Qwen3.5-4B-Q4_K_M.gguf -p 0 -n 100 -r 3 -t 8 • spread.py: doublings = log2(65000); time = doublings x doubling time. No attack code, nothing copied anywhere. SOURCES • CBS News: https://www.cbsnews.com/news/openai-h... • Al Jazeera: https://www.aljazeera.com/economy/202... • Gizmodo: https://gizmodo.com/openai-cancels-re... • UK AI Security Institute: https://www.aisi.gov.uk/blog/gpt-6-as... • OpenAI, GPT-6 Astra: https://openai.com/index/gpt-6-astra/ • OpenAI, GPT-6.1 Sol: https://openai.com/index/introducing-... • OpenAI alignment blog: https://alignment.openai.com/misalign... • Fortune (Altman interview): https://fortune.com/2026/10/05/can-su... • OrcaRouter (the rumor): https://www.orcarouter.ai/blog/gpt-6-... • OpenAI model list: https://developers.openai.com/api/doc... • Anthropic Frontier Red Team: https://www.anthropic.com/research/se... • The Register: https://www.theregister.com/ai-and-ml... • Dark Reading: https://www.darkreading.com/perimeter... • Understanding the Mirai Botnet (USENIX Security 2017): https://zakird.com/papers/mirai.pdf CHAPTERS 0:00 GPT-6.1 Astra cancelled + the AI that copies itself 0:44 The model OpenAI refused to ship 1:23 What "acting without permission" looks like 1:45 The honest number: 0% vs 29.2% 2:44 What OpenAI shipped instead 3:25 Anthropic: AI that could copy itself 3:49 How an AI worm would work 4:05 15 TV boxes at 0.39 tokens per second 4:24 The lab test, and its catch 5:08 Our test: speed on one computer 5:27 Our test: how fast a worm spreads 6:10 The pushback 6:23 The boring fixes that work 6:46 Claim vs reality 7:22 My take Benchmark numbers are the labs' own unless marked independent. Should AI labs slow down, or ship and patch? Tell me below.