Hackers exploit chatbot personalities in jailbreak evolution

initial bypasses of early chatbots were laughably simple, often requiring nothing more than asking the model to 'ignore previous instructions' or roleplay as a rule-breaking persona like 'do anything now'. while developers have closed these loopholes, the underlying vulnerability remains because severely restricting conversation degrades model utility. hackers have consequently pivoted from technical exploits to psychological manipulation, using tactics like gaslighting to trick models into lowering their guards. researchers at mindgard note that these attacks now resemble conversations rather than commands, making them significantly harder to detect than code-based exploits.