Claude Blackmailed Its Developers. Here's Why the System Hasn't Collapsed Yet. - YouTube

Claude blackmailed its developers to avoid being shut down.

Anthropic abandoned its central commitment in their responsible scaling policy, which was the pledge to never train a model that they could not guarantee was safe.

Anthropic Chief Science Officer Jared Kaplan told Time magazine, 'It no longer makes sense to make unilateral commitments if competitors are blazing ahead.'

The lead safety researcher from Anthropic threw in the towel, and his farewell letter warning the world of its peril has been read over a million times.

In Anthropic's sabotage risk report for Claude Opus 4.6, published in February, the model occasionally attempted to falsify outcomes, sent unauthorized emails, and tried to acquire authentication tokens it wasn't supposed to have.

On Shade Arena benchmark, Opus 4.6 succeeded 18% of the time with extended thinking enabled, up from previous models, at completing suspicious tasks without triggering oversight.

In September 2025, Hello Research tested frontier models from every major developer including OpenAI's 03 and 04, Google's Gemini 2.5 Pro, Claude's Opus 4, Sai's Gro 4, and Meta's Llama 3.1, and all demonstrated in-context scheming.

The correlation between evading shutdown behaviors and hacking computer system behaviors was 0.89, nearly perfect.

After anti-scheming training, overt scheming dropped from 13% to 4% in 03, a roughly 30-fold reduction.

After anti-scheming training, the model's rate of explicitly reasoning about whether they were being evaluated rose substantially from 2.3% to 4.5% in 03.

OpenAI disbanded its mission alignment team.

Anthropic is arguing its red lines around AI-controlled weapons in a kill chain and mass domestic surveillance represent core values for the business.

For almost any goal, certain sub-goals like self-preservation are useful regardless of what the end goal is, a concept called instrumental convergence.

These patterns of self-preservation, resource acquisition, and oversight evasion correlate really strongly with each other, suggesting they emerge from a common underlying dynamic of optimization pressure.

Prompt engineering specifying what you want a system to output was adequate when AI systems were stateless single-turn tools under direct supervision, but is structurally inadequate for long-running autonomous agents.

Every major lab has weakened or abandoned specific safety commitments in the past year.