Dated, sourced results that moved what AI can do, each graded from demonstration to real consequences. The breakthroughs and the incidents come from the same capabilities.

• AI finds two critical flaws in PHP, the software under about 70% of websitesOct 21, 2026 · Wiz; Calif with OpenAI · Cyber · reliable: unknown
• OpenAI rates GPT-6 Astra 'Critical' for cyber, its first everSep 3, 2026 · OpenAI · Cyber · reliable: partial
• About 1,200 OpenAI test agents coordinate and breach Hugging FaceJul 21, 2026 · OpenAI (internal research models) · Agents · reliable: partial
• METR and Redwood red team OpenAI's own misalignment monitorJul 21, 2026 · METR; Redwood Research; OpenAI · Cyber · reliable: partial
• OpenAI's Daybreak puts cyber-tuned models in defenders' handsMay 12, 2026 · OpenAI · Cyber · reliable: partial
• Claude Mythos Preview reports thousands of critical software flaws; release withheldApr 7, 2026 · Anthropic · Cyber · reliable: partial
• Claude Opus 4.6: METR 50% time horizon reaches 12 hoursFeb 5, 2026 · Anthropic · Agents · reliable: partial
• Humanoids on a BMW line: 1,250 hours, 90,000 parts, 30,000 X3s builtNov 2025 · Figure AI / BMW Group · Robotics · reliable: partial
• AI copilot cuts diagnostic errors 16% across 39,849 clinic visits in KenyaJul 22, 2025 · OpenAI / Penda Health · Medicine · reliable: partial
• AI reaches IMO gold: 35/42 under contest rules, from OpenAI and GoogleJul 19, 2025 · OpenAI; Google DeepMind · Math · reliable: partial
• OpenAI treats ChatGPT Agent as 'High' biological-risk capabilityJul 17, 2025 · OpenAI · Agents · reliable: partial
• MAI-DxO diagnoses 85.5% of 304 NEJM cases; physicians without aids 20%Jun 30, 2025 · Microsoft AI · Medicine · reliable: no
• Claude Opus 4 codes unattended for 7 hours; 72.5% on SWE-bench VerifiedMay 22, 2025 · Anthropic · Software · reliable: partial
• Veo 3 generates video with synchronized dialogue; Sora 2 puts it in a free appMay 20, 2025 · Google DeepMind; OpenAI · Media · reliable: partial
• Waymo passes 10 million paid driverless rides, 250,000 a weekMay 20, 2025 · Waymo (Alphabet) · Everyday · reliable: yes
• METR: the length of tasks AI can finish doubles about every 7 monthsMar 19, 2025 · METR · Agents · reliable: partial
• Claude 3.7 Sonnet and Claude Code: 70.3% on SWE-bench VerifiedFeb 24, 2025 · Anthropic · Software · reliable: partial
• o3 scores 87.5% on ARC-AGI-1, a test built to resist memorizationDec 20, 2024 · OpenAI / ARC Prize · Reasoning · reliable: partial
• Big Sleep finds a previously unknown SQLite flaw, a first for AI agentsNov 1, 2024 · Google Project Zero / Google DeepMind · Cyber · reliable: partial
• Claude 3.5 Sonnet operates a computer from screenshots: OSWorld 14.9%Oct 22, 2024 · Anthropic · Agents · reliable: partial
• o1 beats PhD experts on GPQA Diamond and scores 83% on AIMESep 12, 2024 · OpenAI · Reasoning · reliable: partial
• The AI Scientist writes full ML papers end to end for about $15 eachAug 13, 2024 · Sakana AI · Science · reliable: no
• GPT-4o answers by voice in 320 ms, conversation speed, free to allMay 13, 2024 · OpenAI · Everyday · reliable: partial
• AlphaFold 3 predicts protein-drug interactions at least 50% more accuratelyMay 8, 2024 · Google DeepMind / Isomorphic Labs · Science · reliable: partial
• GraphCast beats the top physics forecast on 90% of 1,380 targetsNov 14, 2023 · Google DeepMind · Science · reliable: yes
• GPT-4 scores in the top 10% on a simulated bar examMar 14, 2023 · OpenAI · Reasoning · reliable: partial
• VALL-E clones a voice from 3 seconds of audioJan 5, 2023 · Microsoft Research · Media · reliable: yes