Early anthropic users of the Claude Mythos Preview model sign confidentiality agreements to prevent sharing of findings, According to the Wall Street Journalbut the Journal says that all changed last week.
The most important thing to remember about the Claude Mythos Preview so far, is claimed to be the world’s scariest AI modelhas been his privacy. To use it, you must be one of the VIPs allowed to participate Glasswing project— It is said to be a very select group of about 50 companies and organizations.
If you there is One of the Claude Mythos Preview testers involved in Project Glasswing says you’re supposed to use a model to find security vulnerabilities, and the original idea was that participants had a crushing responsibility to keep things secret—as if the fate of the world depended on secrecy.
But according to the Journal, Democratic Representative Josh Gottheimer wrote a letter to Anthropic complaining about it. “No entity should be contractually restricted from notifying others, coordinating mitigation measures, or informing relevant and trusted stakeholders of immediate cyber risks,” Gottheimer writes.
The magazine’s report, published Monday, sounds like Anthropic is struggling to find a basis for what can be done with the results from the Mythos Preview. An anonymous Anthropic spokesperson told the Journal that “Privacy protection was originally something that partners required and was built into the agreements that partners signed,” but added that Glasswing “has matured” and user agreements have “evolved to enable the broad sharing of key data,” including beyond Project Glasswing.
Another event that happened a week ago was the announcement of a similar program called DawnFrom Anthropic’s main competitor, OpenAI. Daybreak was less secretive than Project Glasswing after the jump, allowing anyone to fill out a short form and request a codebase scan with OpenAI’s latest cybersecurity model. CEO Sam Altman Posted in X said he wants to work with “as many companies as possible now”.
It looks like companies have already started talking openly about what the Mythos Preview showed them. For example, I didn’t pay attention to Grant Bourzikas, Cloudflare’s chief security officer. blog post on Monday about what he and his company found while doing the Mythos Preview. This is an informative post that describes Mythos Preview as similar to other bug-finding LLMs, but adds: “What changes with Mythos Preview is that a model can now take these low-severity bugs (traditionally invisible when backlogged) and combine them into a single, more serious exploit.”
But there is an interesting coda at the end of the post. Bourzikas promises to share additional findings with clients soon, and says, “If your team is doing similar work and wants to compare notes, please contact us,” and then provides an email address.
So it sounds like the shroud of secrecy around the Claude Mythos Preview has been lifted ever so slightly. The folks at Anthropic will feel that their model has lost some of its mystique, but an air of mystery surrounding LLM is not something that can last forever.
