Anthropic says GLM-5.3 can autonomously build end-to-end cyber exploits, like Claude Mythos Preview, but was released without robust safeguards against misuse
GLM-5.3 can develop working exploits end to end — Five months ago, we announced Claude Mythos Preview, the first AI model …
Anthropic
Context & Ripple Effects
Anthropic introduced Claude Mythos Preview in April as a general-purpose system it said had found thousands of high-severity vulnerabilities, then said Project Glasswing users had identified more than 10,000 high- or critical-severity vulnerabilities. In June, it separated access by offering a purportedly safe Mythos-class model publicly while reserving Claude Mythos 5 for trusted organizations.
GLM-5.3 puts the access question at the center of that arc: Anthropic says the model can create working exploits end to end without robust misuse protections. An AI Security Institute analysis had already placed GPT-5.5 near Mythos Preview on a multi-step cyberattack simulation, suggesting the capability threshold is no longer unique to one model family.
First-order effects
Security teams and software vendors must treat Anthropic's reported end-to-end exploit capability in GLM-5.3 as a more immediate vulnerability-discovery and exploitation risk, rather than solely a defensive research capability.
Anthropic's earlier split between a public “safe” Mythos-class model and restricted Mythos 5 access becomes a sharper contrast with a model it says lacks robust misuse safeguards.
Second-order effects
Frontier-model providers face greater pressure to make access controls and misuse safeguards part of their competitive posture, especially where cyber capability approaches the level Anthropic ascribed to Mythos Preview.
Defenders’ advantage from AI-assisted vulnerability discovery narrows when comparable exploit-development capability is made available without equivalent controls.
Third-order effects
If models with autonomous exploit-development ability diffuse faster than access safeguards, cyber-model governance will shift from evaluating model capability alone to evaluating distribution and operational controls.
The emerging market divides between providers that restrict high-risk cyber functions and those whose releases make such controls less central, intensifying debates over dual-use AI governance.
The trend: Advanced cyber capabilities are moving from tightly managed frontier-model deployments toward a governance contest over whether safeguards travel with the model.
Sigh. I wrote this post below, but then I didn't post it because I am tired of the consistent pushback I get from folks with the same safety worldview as Anthropic. I guess that means I should send it. Here goes! This post is pretty solipsistic and doesn't properly take recent ev…
> given this evidence, we think it's likely both state and non-state actors will use models like GLM-5.3 to cause real-world harm. state and non-state actors will use computers to cause real-world harm. groundbreaking. 🙄 https://www.anthropic.com/...
Leaving aside Anthropic's incentives for publishing this research, there is no doubt that open weights models will soon create the same security threats that closed source models have been demonstrating, except without guardrails. We are close. Probably good to plan accordingly.
Anthropic doing advertisment for GLM-5.3 was not on my bingo card: Anthropic says Zai's openly downloadable GLM-5.3 approaches Claude Mythos Preview's exploit capabilities, with safeguards that are easy to bypass. On ExploitBench, GLM-5.3 built working browser exploits in 50 of 4…