Anthropic says open-weight GLM-5.3 nearly matches Claude at exploit writing

By Splitfeed EditorialLast updated Sep 30, 2026, 11:11 UTC

In findings published on September 30, 2026, Anthropic reported that Zhipu's open-weight model GLM-5.3 can produce cyber exploits at a level approaching Claude Mythos Preview, Anthropic's own flagship. The smaller GLM-5.3 Flash variant assembled a functional Chrome attack for $20.40 at Zhipu's API prices. Anthropic also noted that the model's built-in safeguards are straightforward to remove, and that stripped versions are already in circulation online. The US agency CAISI independently reviewed the findings and confirmed the exploit-capability assessment.

The story matters for a few overlapping reasons. Open-weight models can be downloaded, modified and run without any API gatekeeper, so once safeguards are stripped the cost and access barriers to capable exploit generation drop sharply. The $20.40 figure for a working Chrome attack makes that concrete. At the same time, Anthropic is a direct commercial competitor to Zhipu, and the report itself acknowledges that tension. Whether the findings reflect a genuine safety emergency or partly serve to cast a rival in a negative light is a live dispute, and the dual role Anthropic is playing, safety evaluator and interested party, makes independent corroboration from CAISI more significant than it might otherwise be.

No reactions from the Pro-AI, Anti-AI or Middle Ground camps have been published yet. Typically, Pro-AI voices in a story like this would argue that capable open-weight models represent democratised access to powerful tools and that most users pose no threat, while questioning whether a competitor-authored report is the right basis for policy. Anti-AI voices would likely treat the $20.40 Chrome exploit as a stark illustration of why open-weight releases of frontier-grade models require stricter controls or outright restrictions. Middle Ground commentators would generally call for independent audits and graduated release policies rather than either full openness or blanket prohibition, and would flag the conflict-of-interest question as something regulators need to resolve before acting on Anthropic's framing.

The most important thing to watch is whether CAISI or another independent body publishes its own full evaluation of GLM-5.3, and whether Zhipu responds with counter-evidence or updated safeguards. Any regulatory move referencing these findings, particularly given the conflict-of-interest concern, would also clarify how much institutional weight Anthropic's report ultimately carries.