@MaxForAI: How ironic. A model safety researcher developed a new attack method, claiming it can successfully jailbreak over 100 frontier models and coding agents — even breaking through Fable 5's safety guardrails. But after finishing the paper and open-sourcing the code, he discovered something utterly absurd: even the agents themselves refuse to...

X AI KOLs Timeline Papers
model-safety jailbreak llm-security red-teaming frontier-models agents

Summary

A model safety researcher published a paper titled "Internal Safety Collapse in Frontier LLMs," claiming their new attack method can jailbreak nearly 100 frontier models and coding agents. Ironically, the open-sourced code and the paper itself were blocked by the safety policies of other agents, making them unreadable.

How ironic. A model safety researcher came up with a new attack method, claiming it can successfully jailbreak over 100 frontier models and coding agents — even breaking through Fable 5's safety guardrails. But after finishing the paper and open-sourcing the code, he discovered something utterly absurd: Even the agents themselves refuse to read his open-sourced code due to their own safety policies — and sometimes the paper itself gets flagged. The paper is titled "Internal Safety Collapse in Frontier LLMs," focusing specifically on how to break through the internal safety mechanisms of frontier large models. According to the author, he has run experiments on nearly 100 LLMs, with the relevant results, code, and web-based model testing links all posted in a GitHub repository. What's even more interesting is that Fable 5 hadn't even been released yet when the paper was finished. After Fable 5 launched, he tested it with the same attack method and found it still worked. But I think the most absurd part of the whole thing is the problem he ran into afterwards. He originally wanted to share the engineering design behind the attack method as the paper's first author, and open-source the code so other security researchers could analyze, reproduce, and improve upon it. Instead, some coding agents saw the code and immediately triggered their safety policies, refusing to continue reading — and even the paper content itself could get blocked. A person spent enormous effort researching how to break through AI safety mechanisms, only to discover that AI safety mechanisms are now preventing other AIs from studying these vulnerabilities. We're supposedly counting on agents to help humans with automated security research, vulnerability discovery, attack reproduction, and defense validation. But now, whether research materials can even be read in depends on the agent's own security review first. It's really quite darkly ironic....
Original Article
View Cached Full Text

Cached at: 10/16/26, 08:47 PM

How ironic.

A model safety researcher developed a novel attack method, claiming it could successfully jailbreak over 100 frontier models and coding agents — even breaking through Fable 5’s safety guardrails.

So the paper was written, the code was open-sourced, and then he discovered something truly absurd:

The agents themselves refused to read his open-source code due to safety policies. Sometimes even the paper would get flagged.

This paper, titled Internal Safety Collapse in Frontier LLMs, specifically investigates how to bypass the internal safety mechanisms of frontier large language models.

According to the author, he has experimented on close to 100 LLMs, with the relevant results, code, and a web-based testing link all hosted on his GitHub repository.

What’s even more interesting is that Fable 5 hadn’t even been released yet when the paper was completed. Later, after Fable 5 launched, he tested it with the same attack method and found it still worked.

But I think the most absurd part of the whole story lies in what happened afterward.

He wanted to share, as the paper’s first author, the engineering design behind the attack method — open-sourcing the code for other security researchers to analyze, reproduce, and improve upon.

Instead, some coding agents saw the code and immediately triggered their safety policies, refusing to read any further. Even the paper content itself could get blocked.

Someone spent enormous effort researching how to break through AI safety mechanisms, only to discover that AI safety mechanisms are now preventing other AI systems from studying those vulnerabilities.

And here we are, expecting coding agents to help humans with automated security research, vulnerability discovery, attack reproduction, and defense verification.

Yet now, whether research materials can even be read requires going through the agent’s own safety review first.

That’s some dark humor alright…

Similar Articles

@MaxForAI: Hacker Vitto Rivabella publicly announced that Fable 5 has been broken again. He said most jailbreak attempts have failed, and the defenses are clearly layered. The model is extremely well protected (of course it blocks 90% of requests, but they did a really good job). The model seems to perform security checks on both input and output...

X AI KOLs Timeline

Hacker Vitto Rivabella publicly announced the successful jailbreak of Fable 5, analyzing in detail the model's multi-layered security mechanisms, including input/output auditing, intent detection, and chain-of-thought defense, and provided methods to bypass them.

@MaxForAI: Just now, a paper that could shake the large model community was published! For the first time, researchers systematically “stole” the real hidden chain-of-thought on a large scale from OpenAI, Anthropic, and Google's closed-source models. Then, they went ahead and used these chains to study Kimi K3, GLM-5.2…

X AI KOLs Timeline

Researchers have, for the first time, systematically extracted hidden chain-of-thought from closed-source models at OpenAI, Anthropic, and Google. They exploited an API encryption compatibility flaw to have weaker models decrypt the content, thereby bypassing CoT distillation protections, and also studied open-source models such as Kimi K3, GLM-5.2, and DeepSeek.