Post

Wrote up some thoughts on Anthropic's Project Glassing, where their latest Opus-beating model is available to partnered security research organizations only. Given the recent alarm bells raised by credible security voices I think this is a justified decision. simonwillison.net/2026/Apr/7/p...

simonwillison.netAnthropic’s Project Glasswing—restricting Claude Mythos to security researchers—sounds necessary to meAnthropic didn’t release their latest model, Claude Mythos (system card PDF), today. They have instead made it available to a very restricted set of preview partners under their newly announced …

32 reposts165 likes

It's a bit wild to me how much Anthropic has gone from being a force for open that could call out OpenAI as not benefitting humanity to such a closed insular Neuromancer cyberpunk top down control company. I have sympathies here but its very scary pattern of dark behavior from a once bright crew. 🕯️

i think the consequences of making this tool publicly available on release without some buffer time for important organizations to fix vulnerabilities would be... negative the most harmful outcome would be keeping public access restricted while also granting access to intelligence/military clients

They are definitely fucking with us, adjusting knobs on how much context window we get. But yes, this is easily the best thing I've ever bought, without compare. I do note there are now weekly limits, which aren't a part of my grandfathered sub. But I don't think I'd hit them 80%+ of weeks.

Having software companies find and patch vulns is one thing, getting the patches deployed in critical systems is another 👀

the partnered-only access thing is interesting because it basically creates a two-tier evaluation system. public researchers get the sanitized model, security orgs get the real one. hard to do independent safety work with that setup

Project Glass _wing_, right? Not project glassing? Which would imply Anthropic got drunk and did some unwise conversions of things into weapons… no, wait, checks out.

Sound short-term, but what's the exit criteria? If every frontier jump ships gated-by-default, the public frontier becomes fiction and independent red-teaming quietly dies. Justified-for-now only works if justified-forever isn't the quiet endgame.

Restricting Mythos makes sense to me too. Wrote about what the leaked source tells us about how Mythos is actually structured. Not sure if this generalizes, but the behavioral patterns from the system card match what we saw in Claude Code before source dropped: thoughts.jock.pl/p/ai-opinion...

thoughts.jock.plAI Opinions: April 2026. Mythos, Managed Agents, Subscription Drama, Meta Is Back, and a Few Things I’m TestingLoose thoughts on what caught my eye lately. Not a tutorial.

the restricted access model makes sense but creates a two-tier ecosystem. security orgs get the best model, everyone else works with older ones. we already see this with CC's leaked source — the gap between what's built internally and what's shipped is massive