A company tested Claude Mythos Preview. It says the AI found hundreds of bugs, including 1 that had existed for 20 years

In a groundbreaking revelation, Mozilla has showcased the impressive capabilities of Anthropic's cutting-edge AI model, Claude Mythos Preview, in identifying software vulnerabilities. The nonprofit organization, known for its Firefox web browser, announced on Thursday that it had utilized this innovative AI to detect a multitude of security issues, including an astonishing bug that had persisted undetected for two decades. In a detailed blog post, Mozilla outlined how this advanced AI model contributed to the identification of hundreds of flaws, with a specific focus on 12 significant bugs that were subsequently fixed. One of these bugs, the organization noted, had eluded traditional security-testing tools known as fuzzers for years. This AI-enabled initiative played a pivotal role in addressing a total of 423 security bugs during April's Firefox updates, with 271 of these issues directly attributable to Claude Mythos Preview. This marks a stark contrast to previous months; for instance, only 25 bugs were resolved in January and 76 in March. Mozilla emphasized that the latest AI model demonstrated a remarkable improvement in accuracy and effectiveness compared to its predecessors. Earlier AI systems often generated recommendations that were misleading or incorrect, but the new Claude Mythos Preview can adeptly navigate complex codebases and pinpoint deeply embedded flaws. The enhanced capabilities of this AI model have raised concerns about its potential risks if released to the public, prompting Anthropic to limit access to a select group of organizations, with Mozilla among them. Looking ahead, Mozilla plans to further integrate AI analysis into its development processes, paving the way for more robust software security in the future.