Anthropic Eases Secrecy on Claude Mythos for 50 Testers After Congressional Complaint
Updated
Updated · Gizmodo · May 19
Anthropic Eases Secrecy on Claude Mythos for 50 Testers After Congressional Complaint
1 articles · Updated · Gizmodo · May 19
Anthropic changed Project Glasswing agreements last week to let Claude Mythos Preview findings be shared more broadly, loosening confidentiality that had bound a small group of testers.
Rep. Josh Gottheimer pushed the issue in a letter arguing no company should be contractually blocked from warning others, coordinating mitigations or informing trusted stakeholders about urgent cyber risks.
Anthropic told the Wall Street Journal the secrecy terms were initially requested by partners, but said the program has matured and the agreements now permit wider disclosure beyond Glasswing.
About 50 companies and organizations were originally allowed into the highly restricted testing effort, which uses Mythos Preview to hunt security flaws and assess how the model can chain minor bugs into larger exploits.
The shift also brings Anthropic closer to OpenAI's more open Daybreak program, announced a week earlier, as Cloudflare has already begun discussing Mythos findings publicly and inviting peers to compare notes.
Anthropic's 'scariest' AI chains minor bugs into major threats. What does this mean for global critical infrastructure security?
As AI finds critical software flaws in minutes, can human security teams possibly keep up with the coming wave of exploits?
Claude Mythos Preview Unveiled: Record-Setting Vulnerability Detection, Dual-Use Dangers, and the Race for AI Regulation
Overview
Claude Mythos Preview marks a major advancement in artificial intelligence, delivering a step change in performance for cybersecurity, reasoning, and coding. Its AI-assisted analysis quickly uncovered 22 Firefox vulnerabilities, with 14 rated as high severity, showing its power to transform defensive cybersecurity for critical infrastructure. However, these unprecedented capabilities come with significant risks, as the model can be used for both defense and offense, raising national security concerns. This dual-use nature highlights the urgent need for democratic countries to maintain a technological lead and develop strong governance to manage both the opportunities and dangers of advanced AI.