🧠 Food for thought

Implications, context, and why it matters.

Anthropic’s new model finds and exploits old software bugs

• Anthropic’s Mythos model has raised concern because internal tests found that it could identify and exploit serious software flaws on its own, including some that had stayed hidden for decades 1. For example, it found a now-patched 27-year-old bug in the OpenBSD operating system, a security-focused open-source software platform 1.
• On a Firefox benchmark, Mythos built working exploits 181 times, while its predecessor Opus 4.6 succeeded twice in several hundred attempts 1.
• Anthropic has not released Mythos to the public. It has given access to a small group of partners and vetted organizations through Project Glasswing, an Anthropic program for testing and fixing vulnerabilities 2. That group includes Microsoft and JPMorganChase 2. US authorities and the Bank of England, the UK’s central bank, have also discussed the model, while regulators in Singapore and South Korea are watching more closely 3.

The narrow Mythos rollout raises questions about access and readiness

• The limited rollout gives some US banks, including JPMorgan and Goldman Sachs, early access to test defenses 4. European financial institutions largely have not received that access so far 4. Some US banks that lack access worry about a competitive disadvantage 4.
• Anthropic and others in the field warn that AI can shrink the time needed to find weaknesses and turn them into exploits 2. Work that once demanded skilled manual effort could happen much faster, which raises pressure on banks and other essential industries to strengthen defenses 2. South Korean officials have discussed ideas such as an AI-based real-time defense system 3.
• Some describe Mythos as a “canary in the coal mine,” an early warning sign that more models with similar abilities are likely to appear 3.