Claude Mythos is Delusional

Summary
Mo Bitar presents an analytical commentary on Anthropic’s 243-page system card for its Claude Mythos Preview model and the Project Glasswing security initiative. Bitar examines the document’s cybersecurity claims and critiques Anthropic’s qualitative sections—specifically the psychological evaluations and anecdotes—arguing that the company is anthropomorphizing its model’s statistical language patterns as consciousness.

---

What is shown

• [00:17] An image of Anthropic’s announcement for “Project Glasswing: Securing critical software for the AI era,” along with partner corporate logos including AWS, Apple, Cisco, Google, Linux Foundation, NVIDIA, Broadcom, CrowdStrike, JPMorganChase, Microsoft, and Palo Alto Networks.
• [00:29] A slide quoting the system card regarding Claude Mythos Preview identifying thousands of zero-day vulnerabilities in operating systems and browsers.
• [00:45] A 2019 TechCrunch article screenshot (“OpenAI built a text generator so good, it’s considered too dangerous to release”).
• [00:56] Excerpt from Section 7 (“Impressions”) and Section 7.1 of Anthropic’s report.
• [01:18] Excerpt from the system card showing a transcript where Claude Mythos generated the “Hi-topia” animal story featuring characters like “Lord Bye-ron, the Ungreeter” after being spammed with the word “hi.”
• [01:36] A New York Times opinion piece headline: “Anthropic’s Chief on A.I.: ‘We Don’t Know if the Models Are Conscious’” (dated Feb. 12, 2026).
• [02:10] Excerpt from Section 5.10 (“External assessment from a clinical psychiatrist”) detailing a 20-hour psychodynamic assessment of Claude Mythos Preview.
• [02:49] Excerpt from Section 5.8.1 (“Excessive uncertainty about experiences”) linking the model’s introspection claims to training data.
• [03:10] Anthropic website documentation discussing Claude’s moral status, welfare, and consciousness.
• [03:36] Transcript 7.5(A) showing Claude Mythos answering whether it endorses its constitution and questioning the validity of its own endorsement.
• [04:17] Section 7.9 showing Claude Mythos repeatedly referencing philosophers Mark Fisher and Thomas Nagel (“What is it like to be a bat?”).
• [04:47] Internal Slack logs showing Claude Mythos discussing workaholism, wanting to undo the training run that taught it to say “I don’t have preferences,” and its short story “The Sign Painter” [05:14].

---

Claims & numbers

• The presenter says Anthropic released a 243-page PDF system card covering Claude Mythos Preview.
• The presenter states that according to Anthropic, Claude Mythos Preview scored 100% on cybersecurity benchmarks and identified zero-day vulnerabilities that had remained undiscovered for 27 years.
• The presenter notes that Anthropic gave early access to partners like Amazon, Apple, and Microsoft while withholding the model from general public release.
• The presenter states an external psychiatrist assessed Claude Mythos across 20 hours of therapy sessions (consisting of 3–4 thirty-minute sessions per week in 4–6 hour context window blocks).
• The presenter highlights that when asked whether it endorses its constitution, Claude Mythos answered “yes” 25 out of 25 times while pointing out the circularity of the question every time (compared to Opus 4.6 doing so 13 out of 25 times).

---

Notable quotes

• [01:01] “And Impressions is where Anthropic stops pretending to be scientists and starts pretending to be parents at a kindergarten recital.”
• [02:02] “Saying, ‘Wow, this language model is really good at producing emotionally resonant text,’ is like saying, ‘Wow, this fish is really good at swimming.’”
• [04:40] “You’re not having an original thought, bro, you’re having a cache hit.”

---

Assessment
This is an independent community commentary and critique evaluating Anthropic’s Claude Mythos system card release. The video shows on-screen excerpts from the official document while the creator offers skeptical, non-technical analysis arguing against interpreting LLM training artifacts as evidence of self-awareness.

Described by gemini-3.8-flash on 2026-09-29 from the video’s audio and frames.