According to Sensics Beating monitoring, Epoch AI released a research report that systematically evaluated the network security capabilities of Anthropic's Claude Mythos series models. Researchers built the "Cybersecurity-Enabled Capability Index" (Cyber-ECI) based on multiple security benchmark tests, and the results showed that Mythos Preview significantly outperformed in vulnerability exploit development capability, leading by approximately 7 months ahead of the linear development trend and also surpassing OpenAI's GPT-5.5 by around 3 months.
Previously, some scholars questioned whether Mythos Preview's performance was on par with GPT-5.5. The report pointed out that the divergence stemmed from early evaluations using a lower-performing internal version, rather than the April version subsequently opened to Glasswing partners, and some test sets were nearing saturation, making it difficult to distinguish the difference. Actual security tests showed that Mythos Preview could frequently autonomously generate arbitrary code execution exploits, a feat early models could hardly achieve.
However, there is a lack of strong evidence in terms of Vulnerability Discovery. Despite a surge in high-severity and critical vulnerability reports (CVEs) from 21 renowned institutions in April and May after Mythos Preview's release, increasing by 142% and 262% respectively compared to the 2025 baseline, the data surge might mainly be attributed to funding. Anthropic provided Glasswing with up to $100 million in API credits, attracting numerous security personnel to conduct concurrent searches in a short period.
For example, in the highly anticipated curl codebase test, Mythos Preview only identified one low-severity vulnerability along with four false positives, showing a performance that did not exceed traditional auditing tools. The startup AISLE also confirmed that some open-source lightweight models could identify the same vulnerabilities. Epoch AI believes that Mythos's actual advantage lies in its extremely low false positive rate and more accurate severity level determination. Compared to directly discovering new vulnerabilities, a more precise determination can significantly reduce the defensive side's manual review time and costs.
