Latest Anthropic AI model finds cracks in software defences
The Mythos Paradox: Anthropic Unveils AI Too Dangerous for Public Release
In a move that highlights the double-edged sword of advanced artificial intelligence, Anthropic has announced the development of its latest model, Claude Mythos. However, unlike previous releases, this model is staying behind locked doors.
The reason? Mythos has demonstrated a terrifyingly "keen adeptness" at exposing software vulnerabilities that have eluded human experts and traditional security tools for decades.
Unearthing the "Undetectable"
During internal testing, Mythos pinpointed thousands of "zero-day" vulnerabilities—security flaws for which no patch or fix currently exists. The model’s reasoning capabilities allowed it to identify weaknesses that were often too subtle for standard automated testing.
Key findings from the Mythos trials include:
- A 27-Year-Old Bug: The AI uncovered a vulnerability in commonly used software that had remained hidden since 1999.
- Defeating Mass Testing: Mythos found a critical flaw in video software that had already been put through more than 5 million automated tests by its creators without detection.
- Exploit Chaining: Anthropic warns that the model can surpass all but the most skilled human hackers by "chaining" multiple minor vulnerabilities together into a single, devastating attack.
Project Glasswing: Arming the Defenders
Recognizing that these capabilities could cause "severe fallout" for national security if leaked, Anthropic has launched Project Glasswing. This initiative aims to use Mythos as a defensive weapon, allowing security specialists to "arm themselves ahead of time."
Anthropic is providing approximately $100 million in computing resources to the project, which includes a coalition of over 40 organizations:
- Tech Giants: Amazon, Apple, Microsoft, and Google (via Vertex AI).
- Cybersecurity Leaders: CrowdStrike and Palo Alto Networks.
- Infrastructure Partners: Cisco, Broadcom, and the Linux Foundation.
"The window between a vulnerability being discovered and being exploited by an adversary has collapsed," said Elia Zaitsev, CTO of CrowdStrike. "What once took months now happens in minutes with AI."
The AI Safety & Political Tightrope
The unveiling of Mythos comes at a turbulent time for Anthropic. In February 2026, the White House issued a directive to terminate all federal contracts with the startup following a dispute over safety guardrails. While that order is currently on hold due to a federal court challenge, Anthropic continues to brief the U.S. government on Mythos's capabilities.
| Model | Status | Primary Focus |
|---|---|---|
| Claude Opus 4.6 | Available | General purpose high-end reasoning |
| Claude Mythos | Gated / Restricted | Advanced coding & cybersecurity research |
A New Era of "Agentic" Threats
The core concern shared by researchers is the rise of agentic AI—models that can work autonomously for extended periods to find and exploit bugs. As these capabilities inevitably proliferate, Project Glasswing represents a desperate race to patch the world's most critical infrastructure before the "Mythos-class" tools fall into the wrong hands.
For now, Anthropic maintains that Mythos will remain a restricted research preview, accessible only to those tasked with reinforcing the digital walls of the modern world.