Why security pros are abandoning frontier models
anthropic and openai's cybersecurity guardrails are forcing offensive researchers to use unregulated open-source models to identify vulnerabilities.
ai labs including anthropic and openai maintain strict vetted programs and guardrails to prevent malicious use of models like mythos in cyberattacks. however, offensive security researchers report that these limitations block legitimate defensive work, such as confirming code vulnerabilities. because cloud-based models also risk leaking sensitive zero-day data, researchers are increasingly avoiding frontier models. instead, they rely on open-source alternatives without safety filters to perform necessary security testing.