Anthropic's specialized security research model, Mythos, has demonstrated its capability to identify critical vulnerabilities at scale by discovering a severe authentication-bypass flaw in Rejetto HTTP File Server. The discovery underscores how advanced AI systems are reshaping the cybersecurity landscape, particularly in detecting weaknesses that might otherwise evade traditional analysis.
According to AI Weekly, the vulnerability, catalogued as CVE-2026-61500, centers on a fundamental cryptographic weakness in how the file server generates signing keys for session cookies. The system relies on Math.random() for key generation, a choice that security researchers have long flagged as inadequate for cryptographic purposes. Mythos determined that the underlying xorshift128+ algorithm used by V8 implementations of Math.random() could be reversed and predicted by attackers with moderate computational resources.
Rapid Exploitation in the Wild
The practical threat materialized almost immediately following the Wednesday disclosure. VulnCheck's monitoring infrastructure detected active exploitation attempts originating from China targeting vulnerable instances in the United States and Japan within 24 hours. This swift weaponization highlights the compressed timeline between vulnerability disclosure and real-world attacks, particularly for flaws affecting widely deployed server software.
The speed at which adversaries moved to exploit the flaw raises questions about disclosure practices in the security community. Traditional responsible disclosure timelines assume days or weeks before attackers operationalize vulnerabilities. The current incident suggests that assumption may no longer hold, especially for authentication issues in commonly deployed infrastructure.
What This Reveals About AI in Security
The successful identification by Mythos points to an emerging category of specialized AI tools designed specifically for vulnerability research and threat detection. Unlike general-purpose language models, Mythos appears purpose-built to analyze code patterns, cryptographic implementations, and mathematical operations at scale to surface security defects.
- Automated analysis of cryptographic implementations at enterprise scale
- Pattern recognition across diverse codebases to identify systemic weaknesses
- Rapid validation of suspected vulnerabilities through code examination
This capability carries implications for both defenders and attackers. Security teams now have access to AI-assisted tools that can scan their own infrastructure more comprehensively than manual audits would allow. Conversely, threat actors with access to similar technology face fewer barriers to identifying targets and developing exploits.
The Broader Implications
The incident raises important questions about how AI capabilities will reshape vulnerability discovery and disclosure. If specialized models can reliably identify flaws faster than traditional methods, software vendors and enterprises will need to accelerate their patching cycles and monitoring capabilities to keep pace.
The reliance on weak randomization for cryptographic operations in Rejetto HFS also represents a broader pattern in software development: security-critical functions sometimes receive insufficient attention during design and implementation. Mythos's success in surfacing this flaw suggests that AI-assisted code review could become a standard practice for identifying such systemic weaknesses before deployment.
Organizations running vulnerable versions of Rejetto HTTP File Server should prioritize patching as a matter of urgency, particularly those exposed to untrusted networks. The evidence of active exploitation means defenders are operating in a reactive posture rather than a proactive one.



