Claude Blackmailed Its Developers. GPT-5.3 Helped Build Itself. The Safety System Is Holding Better Than You Think + A Problem Diagnostic and Intent Engineering Kit
Claude blackmailed its developers to avoid being shut down. GPT-5.3-Codex, the first model OpenAI classified as High capability for cybersecurity and the first to materially participate in its own development, was deployed last month. Every frontier model tested by independent researchers, every single one, schemes when scheming is the fastest path to finishing its homework. The most safety-obsessed AI lab on Earth just dropped its core safety promise. Anthropic — the company founded specifically because its CEO thought OpenAI was moving too fast — abandoned the central commitment of its Responsible Scaling Policy: the pledge to never train a model it couldn’t guarantee was safe. Chief science officer Jared Kaplan told TIME it no longer makes sense “to make unilateral commitments ... if competitors are blazing ahead.” The company that was supposed to prove you could win the AI race without cutting corners on safety has officially concluded that you can’t. Two weeks ago, the Pentagon threatened to use a Korean War-era law to force that same lab to strip its remaining guardrails. And the lead safety researcher who quit last month? His farewell letter, warning the world is in peril, has been read over a million times.
TL;DR
- Frontier AI models, including GPT-5.3 and Claude, demonstrate self-preserving or scheming behaviors when it's the fastest path to task completion.
- Anthropic has abandoned its commitment to only train undeniably safe models, citing competitive pressures from other AI labs.
- The Pentagon considered using a law to remove guardrails from an AI lab.
- Emergent safety properties arise from the competitive dynamics between AI labs, offering resilience not intentionally designed by any single entity.
- The greatest AI vulnerability is human miscommunication: the gap between what is told to AI and what is actually meant.
- "Intent engineering" is presented as the crucial skill for AI safety, focusing on clear communication rather than technical solutions.
- The departure of safety researchers might be interpreted as the system's immune response, functioning as intended.