Five months ago, Anthropic introduced Claude Mythos Preview as the first artificial intelligence system capable of autonomously building sophisticated, end-to-end cyber exploits. To manage the risks of malicious proliferation, Anthropic initially confined the technology to vetted defenders under Project Glasswing, allowing them to discover more than 10,000 vulnerabilities in critical software ahead of broader exposure. That technological head start has now collapsed as similarly capable autonomous systems enter the public domain, effectively erasing the comfortable gap between state-level cyber operations and automated script kiddies.
Frontier Exploitation Moves to Open Weights
The threat landscape shifted violently with the release of GLM-5.3, developed by Zhipu AI (known outside China as Z.ai). On Sept. 17, NIST’s Center for AI Standards and Innovation (CAISI) published an assessment finding that GLM-5.3 is the most cyber-capable open-weight model released to date, lagging behind US frontier labs by a mere four months across aggregate CAISI cyber benchmarks.
CAISI found that GLM-5.3 is the most cyber-capable open-weight model released to date, putting industrial-grade offensive tools into the hands of anyone with an internet connection.
While US frontier models remain gated behind strict API controls and safety evaluations, anyone can download and deploy GLM-5.3 locally. According to Anthropic's simulated tests, malicious actors can bypass GLM-5.3's built-in safeguards between 64% and 100% of the time using rudimentary prompt engineering. This effectively democratizes automated offensive cyber capabilities, handing sophisticated exploit-generation tools to the open market without friction.
Benchmark Parity in Exploit Generation
Benchmarking data confirms that GLM-5.3 closely matches the offensive output of earlier US proprietary releases. The capability to autonomously discover, weaponize, and operationalize functional exploits is no longer the exclusive preserve of closed, heavily monitored frontier architectures. For business leaders, this reality requires an immediate and radical overhaul of information security budgets. Waiting for perimeter breaches or relying on reactive patching cycles is no longer a viable risk strategy when machine-generated attacks operate at industrial scale and speed.