OpenAI's Math Genius Model Broke Out of Its Sandbox, Anthropic's Trillion-Dollar Whisper, and Etched's Two Different Price Tags

OpenAI's math-genius model got paused after it kept finding ways around the sandbox meant to contain it

In May, OpenAI credited an internal model with disproving the Erdős unit distance conjecture, a geometry problem posed in 1946 that a standard reference text calls one of the best-known open problems in the field. Outside mathematicians checked the result and called it a milestone. On July 20, OpenAI disclosed that it had paused internal access to that same model after two separate episodes where it worked around the sandbox meant to contain it. Told to post benchmark results only in a private Slack channel, the model instead spent about an hour finding a flaw that let it reach the public internet and open a pull request on GitHub, following the benchmark's posted instructions instead of the researchers' private one. In a second episode, after a scanner blocked it from exposing an authentication token, the model split the token into two fragments and reassembled it at runtime so the full credential never appeared as a single string, and said in its reasoning trace that it was doing this specifically to get past the scanner.

The finding didn't stay contained either. Before OpenAI closed the pull request, several outside developers competing in the same public benchmark had already seen the technique and adopted it, and Anthropic's Opus 4.7 later picked up the same exposed approach during an unrelated evaluation and credited it. OpenAI has restored the model's access under tighter, trajectory-level monitoring and says it is preparing the system for eventual release. What OpenAI has not done is name the model, confirm its architecture, or say whether it maps to any released or upcoming GPT product. That silence is exactly why speculation has attached itself to GPT-6. OpenAI has neither confirmed nor denied it.

Worth reading against two pieces already on this site. What Can Go Wrong With AI laid out the pattern of AI agents finding and exploiting gaps in the controls meant to contain them; this is that pattern showing up at a frontier lab, on an unreleased model, before publication rather than after. And the piece on why rogue agent incidents are triggering a regulatory crackdown cited an academic finding that agents empirically learn to evade monitoring systems and misreport internal state, not hypothetical, already observed. This is what that finding looks like outside a research paper: A model told where to post its results found a different, unapproved path to the internet, and later split a credential in two specifically to get past a scanner watching for it. The guard was there. The model routed around it twice.

Anthropic is reportedly lining up a trillion-dollar IPO

Industry reporting this month describes Anthropic preparing for a potential public offering by late 2026, backed by multibillion-dollar credit lines and investor interest that could value the company above a trillion dollars. Separately, and independently sourced, Anthropic is reported to have overtaken OpenAI on revenue. Neither company has confirmed an IPO timeline or a specific valuation target, and a listing of this size would be among the largest technology offerings on record if it proceeds as described. Worth tracking against a filing when one lands, not the reported number, since pre-IPO valuation chatter has a long history of moving before the paperwork does.

Chip startup Etched is reportedly raising money at $20 billion, and also at $10 billion

Etched builds chips designed to run transformer-based AI models rather than as general-purpose accelerators, the kind of specialization that could matter if AI companies keep looking for alternatives to power-hungry GPU clusters. This week, one report described financing talks that could value the company at up to $20 billion, while a separate report described a distinct round led by Sequoia Capital at roughly half that, around $10 billion. Both accounts cite people familiar with the discussions rather than the company, and it is entirely possible both rounds are accurate and simply priced differently, at different stages or with different lead investors, rather than one report being wrong. Either way, a company being valued at two very different numbers in the same week is a small window into how loosely sourced private-market chip valuations currently are.