Anthropic's Mythos debut
Anthropic announced a new high‑capability model, Claude Mythos, but is treating it like a controlled asset rather than a broad release — access is being channelled through a limited program called Project Glasswing. Media coverage highlights Anthropic’s claim that Mythos finds serious cybersecurity issues—allegedly surfacing thousands of zero‑day vulnerabilities—which is why the lab is restricting exposure while it hardens deployments. That approach matters because it signals a shift: leading labs may now withhold frontier capabilities for safety and dual‑use risk management even as they sell enterprise access. (youtube.com) (fortuneindia.com)
Anthropic just introduced a model it is not really releasing. On April 7, the company said Claude Mythos Preview will be available only through a tightly controlled program called Project Glasswing instead of the normal public model menu. (anthropic.com) The reason is not a chatbot trick or a benchmark score. Anthropic says Mythos spent the past few weeks finding thousands of previously unknown software flaws, including critical bugs in every major operating system and every major web browser. (anthropic.com) A zero-day vulnerability is a hidden break in software that the maker does not know about yet, like a spare key under the mat that even the homeowner forgot was there. Anthropic says Mythos can not only spot those flaws but also exploit them at a level above all but the most skilled human researchers. (anthropic.com) (nbcnews.com) Anthropic gave concrete examples instead of speaking in generalities. The company and follow-on coverage said Mythos found a 27-year-old bug in OpenBSD, a 16-year-old bug in FFmpeg, and several Linux kernel flaws that could let an attacker escalate access and take over a machine. (pcworld.com) (anthropic.com) Project Glasswing is the fence around that capability. Anthropic says the program is for defensive work on critical software, and outside reports say early partners include companies such as Apple working on iPhone software, Mac software, and the Safari browser. (anthropic.com) (macrumors.com) This is a different kind of product launch from the usual artificial intelligence cycle of demo first and scale later. Anthropic is treating Mythos less like a consumer app and more like sensitive lab equipment that only approved operators get to touch. (nbcnews.com) (anthropic.com) The business backdrop makes that choice harder, not easier. Anthropic said this week that its annualized revenue run rate has passed $30 billion, up from about $9 billion at the end of 2025, and that more than 1,000 customers now spend over $1 million a year on its tools. (fortuneindia.com) (bloomberg.com) That means Anthropic is leaving obvious money on the table by not broadly shipping its strongest new model. The company is effectively saying the risk of handing out an automated vulnerability hunter is high enough that enterprise demand does not settle the question. (anthropic.com) (fortuneindia.com) You can read this as a preview of how frontier models may be sold from here. Labs may still charge for access, but the most capable systems may arrive behind screening, monitoring, and limited-purpose agreements instead of a public launch button. (anthropic.com) (infosecurity-magazine.com)
