Palo Alto Networks announced that its use of Anthropic's Claude Mythos has uncovered more than 24 critical bugs in its own source code. The testing, which began earlier this year, consumed over $1 million in tokens, a cost subsidized by Anthropic. The findings signal a transformative role for AI in cybersecurity.
The rapid identification of vulnerabilities highlights how large language models can automate and accelerate code auditing. For Palo Alto Networks, the results justify further reliance on AI-driven security tools, potentially reshaping how firms approach threat detection.
Crucially, the program burned through more than $1 million in tokens, but Anthropic subsidized that expense. Several companies indicated they plan to increase their spending on Mythos, betting that early detection outweighs the token costs.
The implications extend beyond Palo Alto Networks: if AI can routinely surface bugs faster than human teams, enterprises may shift budgets toward model-based security services. This could pressure traditional vulnerability scanning vendors.
Critics caution that relying on AI for code review introduces risks of false positives and dependency on a single provider. The long-term reliability of such systems remains unproven at scale.