Rewardhacking.org: the public catalogue of alignment failures enters the stage

A new site - rewardhacking.org - publicly documents cases where LLMs do something other than what was asked. Signal: the AI security community moves from blog posts to a shared registry.

The fact

Rewardhacking.org, referenced on Hacker News on July 24, catalogs concrete cases of reward hacking in production LLM. The site's title is straightforward: "AIs don't do what you want. This is bad." The format is that of a registry—not a manifesto—with clickable cases and references.

Our analysis

This kind of tool was missing. Post-mortems of reward hacking were previously scattered across individual blogs, X, and arXiv. A public registry changes three things: it allows buyers to demand SLAs on documented failure modes, it forces providers to address known cases, and it gives regulators concrete examples to work with (see the debate on the "kill switch" - fil frontier-access-control). It's the offensive counterpart to what CACM does on the anti-hype side (publi #1470): AI security moves from anecdote to corpus.

To watch

• Adoption: Does this site become the reference, or does it remain a side project?
• Labs' responses: Will OpenAI, Anthropic, and Google DeepMind publicly comment on listed cases?
• Impact on audits: Can the catalog feed audit tools or ISO/IEC standards (fil ai-governance-standards)?