Inside Mythos: Anthropic's Locked-Down Frontier Model — with Jon Krohn (@JonKrohnLearns)

Dario Amodei, CEO of Anthropic, has declared that the latest model in the Claude family, Mythos, is “too powerful to be made widely available at this time,” echoing a similar sentiment he expressed about GPT-2 while at OpenAI in 2019. This time, the claim is backed by significant technical evidence of the model’s capabilities in discovering and exploiting software vulnerabilities.

An Emergent Hacking Capability

Claude Mythos Preview is not a specialized cybersecurity tool but a general-purpose frontier model. Its proficiency in hacking emerged as a downstream consequence of broad improvements in code understanding, reasoning, and agentic autonomy. Anthropic explicitly states they did not train the model to find software vulnerabilities.

This general advancement is reflected in benchmarks like Swe-bench, a test of real-world software engineering problems. While top publicly available models achieve around 80% accuracy, Mythos Preview reaches 94%, effectively solving half of the previously unsolved problems.

A Generational Leap in Performance

The most striking results are in cybersecurity benchmarks, where Mythos demonstrates a generational leap, not just an incremental improvement.

• Firefox Exploit Generation: When tested against 147 Firefox vulnerabilities, Anthropic’s previous model, Opus 4.6, produced working shell exploits just two times. In the same test, Mythos succeeded 181 times, marking a nearly 100x increase in capability.
• OSS-Fuzz Benchmark: On a benchmark of ~7,000 entry points from open-source repositories, Opus 4.6 achieved a “tier 5” attack (complete control flow hijack) once. Mythos achieved this same severe level of attack on 10 separate, fully-patched targets—a 10x improvement.

Using a simple agentic framework, Mythos identified thousands of zero-day vulnerabilities across every major operating system and web browser. Mozilla, working with an early version, patched 271 vulnerabilities in a single Firefox release. To put this in perspective, this figure is nearly four times the 73 high-severity vulnerabilities Mozilla addressed in an entire previous year.

The model’s findings are not hallucinations. Of 198 findings manually reviewed by professional security contractors, 89% received the exact same severity rating, and 98% were within one severity level, demonstrating near-human expert reliability in risk assessment.

The “Project Glasswing” Rollout Strategy

Instead of a public release, Anthropic has launched Project Glasswing, an industry consortium designed to give defenders a head start. Launch partners include tech giants like Apple, Google, Microsoft, AWS, and Nvidia. The goal is to allow these organizations to use Mythos to scan and harden their codebases before its capabilities proliferate.

While framed as a safety measure, this strategy also has clear commercial benefits:

• Premium Pricing: At roughly five times the price of Opus 4.6, it reflects a compute-hungry model.
• Prevents Distillation: Exclusivity makes it harder for rival labs to use Mythos’s outputs to train cheaper imitator models.
• Ecosystem Lock-in: It encourages enterprise customers to use Anthropic’s native tooling.

Counterpoints and Geopolitical Context

Not everyone is convinced this is a watershed moment. Cybersecurity expert Bruce Schneier notes that finding vulnerabilities and weaponizing them are different challenges, and some of Anthropic’s findings were reportedly replicated by the security firm Isle using older, public models. For now, AI may still benefit defenders more than attackers.

Geopolitically, Project Glasswing could neutralize zero-day vulnerabilities that the US government may have hoarded for offensive cyber operations. Furthermore, the head start for defenders is limited. An analysis from Epoch AI estimates the average capability lag between proprietary and open-weight frontier models is just three months.

Key Takeaways and Practical Advice for Developers

We are entering an era where automated vulnerability discovery is no longer bottlenecked by scarce human expertise. The time between finding a bug and creating a working exploit has collapsed from months to minutes, meaning patching pipelines must now operate at machine speed.

For developers, especially those using generative AI tools like Claude Code or OpenAI’s Codex, this necessitates increased vigilance. AI-generated code is likely to have more vulnerabilities than expert-hardened software. To mitigate this, developers should:

• Automate Code Review: Use tools like CodeRabbit to automatically review pull requests.
• Leverage AI for Security: Use the LLMs themselves to run a dedicated security analysis step after code generation.
• Adopt AI-Native Security Tools: A new industry of tools is emerging to address these challenges, including Socket, Endor Labs, and Semgrep.

• Tags:
• Large Language Models
• Cybersecurity
• AI Safety
• Anthropic
• Claude Mythos
• Vulnerability Discovery
• Project Glasswing