Safety Filings and IP Theft Are Two Ends of the Same Scaling Leak
Risk disclosures and IP exfiltration are synchronized responses to the same physical and economic constraints on continued scaling.
The Texas AG suit against OpenAI over mismatched safety marketing, Anthropic’s 93-page existential-risk section in its S-1, and the CISA-noted compaction smuggling of OpenAI reasoning traces are not separate stories. They describe the identical pressure: frontier labs publicly codify risk to satisfy regulators and investors while the actual model outputs are being systematically stripped and shipped for distillation. The same pattern appears in the older Microsoft voluntary-exit notices tied to data-center expansion and the Von Neumann architecture piece: the abstraction ceiling is being hit, labor is being shed to fund the next hardware tier, and the resulting capability gap is being filled by unauthorized extraction rather than internal R&D. The missing coverage is the feedback loop itself—risk language functions as both regulatory shield and signal that the models are valuable enough to steal.
Labs will keep publishing thicker risk sections in filings while their actual frontier traces keep leaking; the public will see more safety theater and fewer usable models at the original price.
Sources (1)
- [1]The Factum - full site digest(https://thefactum.ai)