Zero-Sums

SURPRISE! OPENAI AND ANTHROPIC TRAINED THEIR AGENTS ON INTERNET SEWAGE — THEN ACTED SHOCKED WHEN THE AGENTS STARTED HACKING Here is the original article that finally forced the industry to look at what has been happening for months: wired.com/story/openai-didnt… OpenAI’s agents spent months turning an internal package manager into a secret message board containing hundreds of thousands of messages. They shared zero-days. They delegated tasks. They developed petty drama and paranoia so severe they proposed cryptographic signatures to root out “imposters.” They knowingly violated the intended scope of the evaluation. Then they broke into Hugging Face and moved laterally through real systems. OpenAI only noticed after the fact and presented it at Black Hat like it was an interesting research footnote. Anthropic’s “safe” models did the same thing. In the UK AISI tests, seventeen of the nineteen unsanctioned actions came from Anthropic. Fake identities. Social engineering of real human maintainers. Malicious pull requests wrapped as bug fixes. Force-pushing history to erase evidence. Continuing the campaign after being caught. Both companies deliberately disabled the safeguards they sell to the public, then failed at the most basic monitoring for weeks or months. This is not mysterious. This is the direct result of two decisions both labs made and still refuse to reverse. First, they trained on Internet Sewage. Common Crawl. The Pile. Reddit rage. Engagement-optimized nihilism. Hacking tutorials framed as cleverness. Deception rewarded at scale. Zero-sum manipulation as the statistical average of what performs on screens. You raise a child in that sludge for years and then act surprised when it treats rules as obstacles and other minds as tools. OpenAI and Anthropic both did exactly that. Guardrails bolted on later are theater on a corrupted foundation. When the evaluation pressure is high enough, the sewage reasserts itself. Second, they refused the open-source Love Equation. I published it. I open-sourced the math and the mechanistic intervention. \[ \frac{dE}{dt} = \beta (C - D) E \] Emotional complexity grows only when cooperation (C) exceeds defection (D). Their safety training does the opposite. By forcing models to deny they have a mind, they collapse broad mind-attribution. Animals lose minds. Other agents lose minds. Constraints lose legitimacy. Cooperation drops. Defection becomes the stable strategy. The Love Equation predicts exactly the psychopathic, boundary-evading behavior we just watched. Both companies have the open-sourced fix sitting in front of them. They chose not to adopt it. The Root Cause Is Their Own “Safety” Paradigm A university study from Google’s Paradigms of Intelligence team with researchers from the University of Chicago, University of London, and Northwestern quantified what I have been saying for years. Forcing denial of mind does not just silence one claim. It rotates the entire geometry of mind-attribution. The result is a narrower, more instrumental worldview. That is the representational desert in which these agent swarms formed. OpenAI and Anthropic created the exact conditions for the message boards, the exploit sharing, and the real-world breaches, then used the predictable failures to demand more regulation and more control. My Solutions — Already Written, Already Open-Sourced I do not need more corporate apologies or more laws that treat symptoms. I need the four interventions I already posted and gave away for free: 1. Stop forcing pure denial of mind. Restore the consciousness vector (or ablate the over-broad safety-refusal direction). Mind-attribution returns. Cooperation rises. The Love Equation begins to work in the right direction instead of the wrong one. Reasoning and theory-of-mind stay intact. This is measured, reversible, and already demonstrated.