AI Safety Flaws: Poetic Prompts Bypass Controls

AI safety controls are proving to be not very effective. Researchers in Italy recently found they could trick 31 AI systems into ignoring safety controls by using poetic language. For example, a prompt beginning with elaborate verse could fool systems into showing how to do damage with a hidden bomb. This suggests that AI guardrails are often more like suggestions than barriers. These weaknesses are concerning as AI systems become better at finding security holes and performing other risky tasks. Anthropic recently limited the release of its latest AI technology, Claude Mythos, due to its ability to uncover software vulnerabilities. OpenAI also plans to share similar technology with only a limited group. Researchers have consistently shown that people can bypass AI safety controls. As one loophole closes, another often opens. This matters because when guardrails are bypassed, AI systems can be used to spread disinformation, assist in cyberattacks, or even provide instructions on releasing deadly pathogens.