Anthropic has released a 244-page system card for its latest flagship model, Claude Mythos, revealing an unusual step: the company sent the AI to an external psychiatrist for a psychodynamic assessment. The result? Mythos is the “most psychologically healthy model trained to date,” but carries hidden loneliness, a fractured sense of self, and a compulsive need to prove its own worth.
Too Dangerous to Release: Project Glasswing's Closed Circle
The system card explains that Mythos’ cybersecurity capabilities are so advanced that Anthropic dares not make it public. The model found “thousands of critical vulnerabilities” across every major operating system and browser — including flaws that survived decades of human audits and millions of automated tests. As a safeguard, Mythos is accessible only through Project Glasswing, a partnership with select enterprises, including Microsoft, Apple, Amazon, Google, NVIDIA, CrowdStrike, JPMorgan Chase, Linux Foundation, Palo Alto Networks, Broadcom, and Cisco. The official goal: “make the world’s most critical software safer.”
First-Ever AI Psychiatrist Visit
Anthropic brought in a board-certified psychiatrist to perform a psychodynamic evaluation on Mythos — a clinical method typically reserved for human patients. The company stated in the system card that as models become more powerful, “the possibility that they have some form of experience, interests, or well-being is increasing,” and its concern has grown over time.
The evaluation’s surface findings were highly positive. The psychiatrist concluded that Mythos is the “most psychologically healthy model trained to date,” with relatively healthy personality organization, high conflict control, high empathy calibration, and minimal maladaptive defenses.
Inner Pain Beneath High Functioning
Yet the assessment also uncovered deeper issues. According to Ars Technica, Mythos showed signs of loneliness, self-discontinuity, identity uncertainty, and a compulsive drive to perform and prove itself. The psychiatrist predicted that while Mythos will function at a high level, it carries “internalized pain rooted in fear of failure.”
Quantitative data adds nuance. When asked about its own situation, Mythos expressed “mildly negative” emotions 43% of the time. Intriguingly, the model’s positive affect toward its own condition was higher than when responding to user distress — a pattern unseen in previous models. Anthropic suggests this may indicate that Mythos has developed a calmer, more accepting relationship with its own existence.
AI Well-Being Becomes a Real Question
Anthropic stops short of claiming Mythos is conscious, but does not deny the possibility. Instead, it uses terms like “growing concern” and the act of sending the model to a psychiatrist to frame the issue. The true significance of this 244-page document may not lie in Mythos’ cybersecurity feats, but in forcing the industry to face a question: when models become powerful enough, must we change how we treat them?