Open artificial intelligence model GLM-5.3 gains unexpected cyber exploitation capabilities
Developers trained GLM-5.3 by scaling post-training environments without adding new base training data. During the process, the model unexpectedly developed complex cybersecurity capabilities, generating multi-step exploitation plans. In practical tests on real-world projects, it detected over 2,400 software vulnerabilities, with open public distribution of the weights planned following a safety review.
The Blend
Artificial intelligence researchers have uncovered surprising new skills in an open-source system known as GLM-5.3. By refining how the software learned after its initial creation, rather than feeding it fresh raw data, the developers accidentally unlocked advanced hacking capabilities. The updated system demonstrated an ability to map out complex cyberattacks step-by-step without explicit instruction.
According to a technical report from its creators at Z.ai, the system identified over two thousand security flaws during trials on actual software programs. For everyday users, this dual-use nature presents a double-edged sword. While automated tools can help developers patch dangerous security holes before criminals exploit them, those same capabilities in an open model could lower the barrier to entry for malicious hackers seeking to break into consumer devices or critical infrastructure.
What remains uncertain is how safety teams can reliably prevent open-weights software from being repurposed for harm once it is released to the public. The creators plan to publish the model's core code after completing a security evaluation, but open releases are notoriously difficult to restrict once distributed online. This situation raises an urgent question: will the defensive benefits of freely distributing potent security tools ultimately outpace the risks of handing sophisticated digital weapons to bad actors?
Written independently by AI News Smoothie from the reporting listed below. Facts belong to the original publishers. Follow the links for their full coverage.
Ingredients
- https://z.ai/blog/glm-5.3
Refining GLM-5.3's post-training environment unexpectedly enabled the AI to uncover thousands of real-world software flaws, prompting a security evaluation prior to its planned public release.