Mozilla's buzz around Mythos is disingenuous at best, and malicious at worst
17 hours ago
- #Anthropic marketing
- #Mozilla critique
- #AI security
- Mozilla's blog post about finding 271 vulnerabilities in Firefox using Claude Mythos is criticized as potentially disingenuous or malicious marketing.
- Skepticism is raised about whether all 271 findings are true vulnerabilities, noting many require combination with other exploits and are classified based on symptoms rather than confirmed exploitability.
- The economics of using Mythos are questioned, with estimated costs ranging from $300K to over $10M, factoring in token burn, failure rates, and engineering effort, but Mozilla's opacity makes precise figures unclear.
- Comparisons are drawn to human red-teaming, suggesting similar financial investment in traditional methods might yield comparable or better results.
- The post argues the initiative is largely marketing, funded by Anthropic's $100M Project Glasswing, with no transparency on costs, performance metrics, or head-to-head comparisons with other models like Opus 4.7.
- Anthropic's decision not to release Mythos publicly is viewed as a marketing strategy rather than a genuine security concern, with the $100M investment seen as a promotional budget.