An initiative to secure the world's software | Project Glasswing
Anthropic · 2026-04-07 · official · 436,437 views
What's in the video
Description written by Gemini, which watched and listened to the whole video.
Summary
Anthropic presents an official announcement introducing Claude Mythos Preview, a frontier AI model exhibiting advanced cybersecurity capabilities, alongside "Project Glasswing." The video features Anthropic leadership (CEO Dario Amodei, red team lead Logan Graham, researcher Nicholas Carlini) together with security executives from Microsoft, Palo Alto Networks, Cisco, CrowdStrike, and the Linux Foundation discussing defensive AI deployment.
What is shown
- [00:00 - 01:23] Interviews with industry leaders (Jim Zemlin of the Linux Foundation, Elia Zaitsev of CrowdStrike, Igor Tsyganskiy of Microsoft, Lee Klarich of Palo Alto Networks, Anthony Grieco of Cisco) detailing how software bugs permeate critical infrastructure.
- [00:38 - 00:44] Animated grid visual illustrating bug proliferation and exploit propagation across interconnected software systems.
- [01:24 - 01:29] Anthropic model lineup graphic showing Mythos PREVIEW positioned above Opus, Sonnet, and Haiku.
- [02:08 - 02:24] Interview segments explaining multi-step vulnerability chaining.
- [02:56 - 03:06] Title card and logo graphic introducing Project Glasswing.
- [03:36 - 04:26] Anthropic researcher Nicholas Carlini describing vulnerability scanning on open-source codebases, including OpenBSD and Linux.
- [05:43 - 05:48] Closing slate displaying the URL
anthropic.com/glasswing.
Claims & numbers
- Capability origin: Dario Amodei claims the model was not trained specifically for cybersecurity, but gained cyber capability as a side effect of general code training [01:45].
- Human parity: Igor Tsyganskiy claims Claude Mythos is "by and large as good as a professional human at identifying bugs" [01:56].
- Vulnerability chaining: Nicholas Carlini states the model can chain 2, 3, 4, or sometimes 5 vulnerabilities in sequence to execute sophisticated exploits [02:18].
- Autonomy: Logan Graham claims the model can autonomously pursue long-range tasks comparable to what a human security researcher would complete over the course of an entire day [02:30].
- Controlled release: Logan Graham states Anthropic will not release Claude Mythos Preview widely due to dual-use exploit risks [02:47].
- Discovery rate: Nicholas Carlini claims he found more bugs in a couple of weeks using the model than in the rest of his life combined [03:37].
- OpenBSD 27-year vulnerability: Carlini states the model discovered a flaw in OpenBSD that had been present for 27 years, allowing an unauthenticated remote crash via a few packets [03:53].
- Linux privilege escalation: Carlini states the model discovered vulnerabilities in Linux allowing an unprivileged user to elevate to administrator privileges; maintainers have patched the discovered flaws [04:05].
Notable quotes
- "Claude Mythos Preview is a particularly big jump along that point. We haven't trained it specifically to be good at cyber... it's also good at cyber." — Dario Amodei [01:41]
- "I found more bugs in the last couple of weeks than I found in the rest of my life combined." — Nicholas Carlini [03:37]
- "For OpenBSD, we found a bug that's been present for 27 years, where I can send a couple of pieces of data to any OpenBSD server and crash it." — Nicholas Carlini [03:53]
Assessment
This is an official announcement and partner showcase video announcing Claude Mythos Preview and the Project Glasswing defensive initiative. It presents verbal testimonials and post-mortem descriptions of patched vulnerabilities rather than live screen recordings, code walkthroughs, or interactive exploit demonstrations.
Described by gemini-3.8-flash on 2026-09-29 from the video's audio and frames.