An AI state of the union: We’ve passed the inflection point & dark factories are coming
Simon Willison argues November 2025 was the AI coding inflection point and predicts a Challenger-style disaster from unsolved prompt injection.
- November 2025 inflection point: GPT-5.1 and Claude Opus 4.5 crossed threshold from ‘mostly works’ to ‘almost always works’ for coding agents.
- StrongDM spent $10,000/day on tokens running swarm AI testers simulating employees in a fake Slack/Jira environment — no humans reading the code.
- Willison predicts a Challenger-style AI disaster: normalization of deviance around prompt injection means a catastrophic breach is inevitable, just not yet.
- The ‘lethal trifecta’: an agent with (1) private data access, (2) exposure to malicious instructions, and (3) an exfiltration path is an unsolvable security hole at scale.
- 97% prompt injection filter effectiveness is a failing grade — three in a hundred attacks still steal all your data.
- Mid-career engineers most at risk; juniors and seniors both benefit — juniors from onboarding help, seniors from amplified 25-year expertise.
- Willison’s three agentic patterns: red/green TDD phrasing forces agents to write tests first; thin project templates enforce code style; ‘hoarding’ tasks agents can’t yet do to track capability advances.
- OpenClaw went from first line of code (Nov 25) to Super Bowl ad in 3.5 months — proof of massive demand for personal AI assistants despite severe security flaws.
2026-04-02 · Watch on YouTube