Anthropic built a moat around its most powerful AI model yet. When tested, the defenses could not hold.
On April 7, Anthropic announced the Claude Mythos Preview, a model expected to rival leading large language models (LLMs) like ChatGPT and Gemini.
On the same day, Anthropic’s red team — which researches risks associated with Claude — reported that Mythos was “strikingly capable” at cybersecurity tasks.
The researchers had discovered that with the right prompt, Mythos Preview could identify and exploit security flaws in every major browser and operating system, including a 27-year-old OpenBSD bug.
Mythos’ offensive capabilities were not by design. Anthropic says they surfaced as a result of broader improvements in code and autonomy—making it better at both finding and exploiting vulnerabilities.
In the hands of Anthropic’s red team, a tool like Mythos is a powerful asset. In the wrong hands, it becomes something far more dangerous, with consequences extending beyond corporate systems into national security threats.
Enter ‘Project Glasswing’
Following the critical discoveries about the Mythos Preview, Anthropic has chosen not to release the model publicly. Instead, it launched Project Glasswing.
It is a controlled initiative initially limited to 12 technology and cybersecurity firms, including Amazon Web Services, Apple, Microsoft and CrowdStrike, with the goal of using Mythos to help safeguard the world’s most critical software.
The idea is straightforward: participating companies use the model to harden their own systems, and Anthropic publishes findings for the wider industry to benefit from.
The project has since expanded to 40 additional organizations, backed by $100 million in usage credits.
Anthony Grieco, Cisco’s Chief Security Officer, captured the spirit of the program in a blog post writing, “Security has always been a team sport… that is what Project Glasswing represents.”
The gap in the moat
Anthropic’s efforts to contain its Mythos model via Project Glasswing could not prevent unsanctioned access.
According to Bloomberg, a small group of unauthorized users in a private forum gained access to Mythos Preview on the same day the Glasswing project was announced.
The entry point was reportedly a third-party vendor relationship, caused by poor access controls, rather than a failure of the model itself.
While the group did not use the model for hacking, the incident raised concerns about unmonitored access in vendor environments and broader AI security risks.
Raluca Saceanu, CEO of Irish cybersecurity firm Smarttech247, told the BBC that unauthorized use of powerful AI tools risks more than a security incident, but also enables fraud and cyber abuse.
Anthropic says it’s investigating, but has found no evidence of a breach on its systems.
