Last month, the Trump administration forced AI company Anthropic to withdraw its most powerful Mythos 5 and Fable 5 AI models, citing “national security” issues. The government had gotten wind that Fable 5’s internal safety guardrails could be circumvented or “jailbroken,” in major vulnerabilities that spooked officials.
Anthropic itself has long warned that its latest models are adept at spotting cybersecurity vulnerabilities that can be exploited by bad actors. Up until the government clamped down, the company had only given a limited number of businesses and institutions access to Mythos 5, the more powerful and less restrained of the two models, to give them a chance to patch up any gaps in security.
Weeks later, the US government once again gave Anthropic green light to resume its limited roll out of Mythos 5. As of July 1, Fable 5 is once again available to the public, but for a steep off-subscription price, and Mythos 5 is available to around 100 US organizations and government agencies.
However, hackers looking to use AI to commit crimes are rapidly getting new options. As Forbes reports, Beijing-based AI company Z.ai has released an open-weight model, dubbed GLM-5.2, that can perform similar large-scale coding tasks and trade blows with Anthropic’s latest models when it comes to identifying vulnerabilities.
GLM-5.2 can be downloaded by anybody, can be run on virtually any hardware, and unlike Mythos or Fable, there’s no vendor playing the middle man between the AI models and the users, raising the cybersecurity stakes considerably.
Put simply, while these frontier models can aid researchers in patching holes in commonly used software, the can also be abused by hackers to bypass existing defenses.
Security firms Semgrep and Graphistry both found that GLM-5.2 was proficient at identifying software bugs and performing other cybersecurity tasks. “We Have Mythos at Home,” Semgrep titled its benchmarking.
Researchers at Graphistry also suggested Z.ai may have distilled OpenAI’s GPT-5.5 and Anthropic’s Opus 4.8 — a controversial practice that involves using one “teacher” AI to train a “student” model — which could explain how China was able to catch up so quickly, as Axios reports.
Consultants also told the publication that hackers are already trading GLM-5.2 jailbreaks on Russian-language forums, suggesting it’s a very real threat.
“An attacker can run it locally without safety guardrails, fine-tune it against their specific targets, and operate with zero visibility to any provider or defender,” AI cybersecurity firm Armadin CTO and founder Travis Lanham told Axios.
More on Mythos: The White House Suddenly Seems Pretty Terrified of Anthropic