Anthropic has quietly revealed new details about a powerful internal AI system that is already shaking up the cybersecurity world in a massive way. The company shared its first major update on Project Glasswing, a cybersecurity initiative launched earlier this year that basically uses advanced AI to stop future AI-powered cyberattacks before they happen. And according to the latest numbers, the results already look pretty serious.
At the center of the project is an unreleased model called Claude Mythos Preview, which Anthropic says has helped partners uncover more than 10,000 software vulnerabilities in just over a month. Even more surprising is the scale of those discoveries. The company claims many partner organizations individually found hundreds of critical and high-severity bugs using the system, dramatically increasing the speed at which security flaws are being discovered and patched.
One of the biggest examples came from Cloudflare, which reportedly identified around 2,000 vulnerabilities with help from Mythos Preview, including roughly 400 classified as critical or high-risk. Mozilla Foundation also saw major results, saying it discovered and fixed 271 Firefox vulnerabilities using the model. According to reports, that number was almost ten times higher than what it previously managed using older AI systems. The sudden jump in bug detection has become one of the strongest signs yet that AI may completely transform cybersecurity over the next few years.
Anthropic also revealed it scanned around 1,000 open-source software projects internally over recent months. Out of more than 23,000 vulnerabilities found, over 6,000 were reportedly considered high or critical in severity. That’s an enormous amount of potentially dangerous flaws hiding inside software many people use every single day without realizing it.
Anthropic Is Still Refusing To Release The Model Publicly
Despite all the impressive results, Anthropic says it still refuses to release Mythos Preview publicly because the risks are simply too high right now. The company admitted that neither it nor any other AI company has built strong enough safeguards to stop such advanced systems from being misused by cybercriminals or hostile groups. And honestly, that concern feels pretty valid considering how quickly AI-generated attacks have already evolved over the last year alone.
There’s also another reason the report is creating attention across the tech industry. According to recent claims from a security research firm, Mythos Preview may have even helped researchers discover a method to breach macOS, which has long carried a reputation for being one of the more secure consumer operating systems. Anthropic did not directly include that case inside its official report, but the mention alone has already fueled discussion about how powerful the system may actually be behind closed doors.
At the same time, Anthropic appears to be rebuilding stronger ties with governments and major tech companies through the Glasswing initiative. The company confirmed it is already working alongside partners including Amazon Web Services, Apple, Google, NVIDIA, CrowdStrike, Palo Alto Networks, and JPMorgan Chase. The company also hinted that governments, including the United States, may soon gain broader access to the initiative as cybersecurity threats continue growing globally.
Meanwhile, Anthropic itself is entering a pretty major moment financially. Reports from The Wall Street Journal recently suggested the company could become profitable for the first time since launching in 2021. It is reportedly on track to generate more than $10 billion in revenue with over $500 million in operating profit for the quarter ending in June. Still, the company apparently does not expect profitability to last consistently because it plans to continue spending aggressively on computing infrastructure and AI expansion.
What makes this whole situation fascinating is how quickly the AI arms race is now shifting beyond chatbots and image generators. Companies are no longer just competing to build smarter assistants or better search tools. They’re now building AI systems capable of discovering hidden vulnerabilities across the internet faster than human teams ever could. And if tools like Mythos Preview keep improving at this pace, the future of cybersecurity may end up looking completely different within only a few years.
