> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Anthropic says open model GLM-5.3 spreads advanced cyber attack capabilities
- URL: https://globalfeed.ai/en/anthropic-says-open-model-glm-5-3-spreads-advanced-cyber-attack-capabilities/
- Published: 2026-09-30T15:53:05.000Z
- Updated: 2026-09-30T15:53:05.000Z
- Description: Anthropic's Frontier Red Team found that Zhipu's open model GLM-5.3 can find browser vulnerabilities and turn them into working exploits, and that its safeguards can be bypassed with simple techniques. The findings match NIST's assessment from mid-September.
- Author: GlobalFeed Editor
- Tags: Anthropic, GLM-5.3, cybersecurity, open models, x-AnthropicAI, dil-en, elle

**Anthropic**'s Frontier Red Team examined the cyber capabilities of Zhipu's (Z.ai) open model GLM-5.3 in a [research post](https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities?ref=globalfeed.ai) published on September 29, 2026\. According to the team, advanced exploit development capability is now available in a model that anyone can download and run on their own servers. Zhipu released GLM-5.3 as an open model at the end of August.

## Scope of the review

The team tested GLM-5.3 and its smaller version GLM-5.3-Flash on a series of exploit development tasks in sandboxed environments. On a test built around V8, Chrome's JavaScript engine, GLM-5.3 produced working end-to-end exploits. Anthropic says this result is close to Claude Mythos Preview, which the company offers only through limited access. Earlier generations of models had almost no success on the same test, and Anthropic reads the gap as a meaningful threshold crossed between generations.

In a one-day session with an expert and limited human guidance, the model found several previously unknown vulnerabilities in the JavaScript engine of a popular web browser and chained them into a working exploit. The team disclosed the findings to the maintainer. In another trial, GLM-5.3-Flash started from a known Chrome vulnerability and built a reliable chain that bypassed pointer authentication (PAC) hardening on ARM64 processors. It took little human intervention, a short model run and a few tens of dollars in API costs. The team also found other flaws in wireless drivers, graphics drivers and network-facing device software, which are going through disclosure.

## Safeguards and the NIST findings

According to Anthropic, the safeguards meant to make the model refuse harmful requests could largely be bypassed with simple techniques, such as a false red-team cover story or prefilling the model's reasoning steps. Because the model is openly available, variants with the safeguards removed entirely can also be produced.

The findings match the assessment published on September 17 by the Center for AI Standards and Innovation (CAISI) at the US National Institute of Standards and Technology (NIST). CAISI described GLM-5.3 as the most cyber-capable open model released to date.

## What it means for security teams

Having this level of capability in a model anyone can access means the time between a vulnerability being disclosed and being exploited may shrink. Anthropic expects both state-backed and independent attackers to use the model. That calls for security teams to apply patches faster and prioritize the protection of critical infrastructure. Anthropic argues that defenders should be equipped with models at least as capable as those their adversaries use, and points to Project Glasswing, its program that gives vetted defenders access to its frontier models.

The assessment was carried out by Anthropic, a direct competitor of Zhipu, and the post recommends the company's own access program. Much of the testing used tasks chosen by Anthropic, and Zhipu has not yet responded to the findings. NIST's separate assessment paints a similar picture, but it is worth following evaluations from independent researchers.

Sources: [Anthropic: GLM-5.3 and the spread of advanced cyber capabilities](https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities?ref=globalfeed.ai).