TechFlow News, June 24: According to an Associated Press report, a U.S. official stated that during its collaboration with U.S. intelligence agencies for testing, Anthropic’s AI model Mythos identified vulnerabilities in certain highly sensitive U.S. government systems within hours—but this does not mean the model successfully exploited or breached those systems within the same timeframe. The tests are part of the “Glasswing Program,” designed to assess the potential cybersecurity impacts of advanced AI on public safety, national security, and economic systems.
The report also notes escalating disagreements between Anthropic and the Trump administration over AI usage and regulation. Earlier, U.S. authorities requested restrictions on foreign access to Anthropic’s Fable 5 and Mythos 5 models—prompting opposition from the cybersecurity industry. Multiple experts argue that while such models possess strong vulnerability discovery and exploitation capabilities, these capabilities are not unique; overly restrictive measures could undermine U.S. cyber defense capabilities.
