AI agents make new attacks that get 100% on a strong model
Agents that write code made new attacks on AI models, and one attack got 100% success on a model made to stop attacks.
Researchers used the agents Claude Code and Codex to make new attacks on large language models (LLMs), the AI that writes text. The agents had a library of more than 30 methods that persons made before. On Meta-SecAlign-70B, a model made to stop attacks, the new method got 100% success. An automated method from before got 82%.
How it works
The agents work in a loop that the authors call autoresearch. An agent writes a method, does a test with a script, and tries again.
The agents made the methods on other models for one task. The methods also gave good results on prompt injection, an attack that puts new instructions in text that the model reads.
The numbers
- On GPT-OSS-Safeguard-20B, one new method got 80% success on queries about weapons.
- Methods from before got less than 50% on the same test.
- On Meta-SecAlign-70B, the new method got 100% and a method from before got 82%.
What is not known
These results come from the authors and their tests on 2 models. The source gives no tests on other models. The authors write that a defense must get a test with attacks made for that defense. They write that autoresearch must be the minimum test.
Insight: Attack tools now improve themselves, so defense tests must too
Other tools report very high success too. RedDarwin got 100% with only 10 queries, and CKA-Agent got more than 95% on 4 models. This reads as a race: attackers use agents to work at high speed, and persons write defenses more slowly. A defense that one set of tests approved can become out of date. Leaders must tell vendors to test with attackers that adapt.
Sources
Posted
